AI Agents: The Most Valuable Skill You Can Learn in 2026 (Full Course)
Remy Gaskell, a self taught non technical operator, spends two and a half hours sharing his screen and taking the whole agent stack apart: what an agent actually is (chat is question to answer, an agent is goal to result), the observe think act loop every harness runs, and the four levers you can pull on it. His argument is that you do not build agents, you tune them, and that setup should be treated as onboarding an employee through context, tools, and skills. He builds a folder live from empty, adding a CLAUDE.md, a context folder, a self updating memory file, an MCP connection to beehiiv, and a working skill, showing the behavior change at each step. Along the way: the six MCP connections he thinks everyone needs, why progressive disclosure lets one agent hold hundreds of skills, his holding company folder structure, his honest $1,000 monthly AI bill, and why companies go AI native from the bottom up rather than the top down.
Published Jul 22, 20262:24:11 video69 min readAdded Aug 7, 2026Open on YouTube →
At a glance
Remy Gaskell is a self taught, non technical operator who ran an ad agency, taught himself agents by asking Claude how everything works, and now runs the AI agent team for one of the fastest growing ecommerce brands in the world. Over two hours and twenty four minutes on Open Residency with host Mark Mastrandrea, he shares his actual screen and takes the whole system apart: what an agent really is, the three step loop underneath every agent product on the market, and the four levers you pull to make one useful.
The spine of the episode is a single claim. You do not build agents, you tune them. The loop is already built by Anthropic and OpenAI; what you own are the context files, the tools, and the skills you hand it, and those are what turn a generic model into something that works like a trained employee. He builds a folder live on camera, starting from empty, adding a CLAUDE.md, a context folder, a self updating memory file, an MCP connection to beehiiv, and finally a skill, and at each step he shows the agent's behavior changing.
The rest is specifics: the six MCP connections he thinks everyone needs, the anatomy of a skill and why progressive disclosure keeps your context window from filling up, orchestrator skills that call other skills in sequence, the global versus project split he calls a holding company, his honest monthly AI bill, and his argument that companies go AI native from the bottom up rather than the top down.
This page rebuilds the whole episode in its own order and keeps the numbers, the tool names, the demos, and the jokes.
The deep explanation
Stage one was chat. Stage two is agents.
Remy opens with the divide he sees forming. Stage one was chat: ChatGPT, Claude, Gemini. You would be hard pressed to find a founder or business owner not using a chat model by now. Stage two is agents, and the founders and employees who have made the jump are, in his words, "no word of a lie, like five to 10 times more productive in their day."
Mark pushes back that it is probably more than that. Remy takes the middle of his own range and does the arithmetic out loud: seven times more productive means a week of output in a working day, and five to seven weeks of output in a week. Stack that over a month or a year "and it's toast." A massive divide is opening between the people using agents and the people stuck in chat, and you want to be on the agent side.
Mark frames it in operator terms. The metric everyone is racing to push down is opex, labor as a percentage of revenue, and most people are trying to move it with chat while the real multiplier sits with agents. Remy adds the flip side that he returns to at the end of the episode: this is an opportunity for employees, not just a threat. Take the initiative to learn this and you become a 100x employee, one go to market strategist absorbing the work of a ten person team while getting paid like a five person team.
The concrete example of what gets given back is small and mundane, which is the point. Someone on Remy's team asked for passwords and API keys. The old path was open 1Password, find them, create a share link, share it securely, send the Slack message, roughly ten minutes. The new path was telling his agent to go into 1Password, create the share links, and send them on Slack. Thirty seconds. Multiply that across a day and you get your time back for the work that is actually fulfilling, creative, and strategic, the visionary work the model cannot do.
Mark's prediction, stated flatly: there was a company worth a billion dollars run by one person, and he believes one to three person teams will do billions and billions.
What "agent" actually means
The word has been beaten to death. Open Instagram or X and it is nothing but "I built an agent that does this." So Remy gives the definition he uses to cut through it:
Chat is question to answer. An agent is goal to result.
With chat you ask a question, it replies, and then you go do the work. With an agent you hand over a goal, it breaks that into a plan, executes, and hands you a finished product. The word itself comes from agency, meaning the ability to act. People hear "agent" and think real estate agent or travel agent, some sci fi guy living his own life and doing errands for you. In reality the term came from agentic AI, meaning AI that can do things, and people shortened it.
His preferred analogy: a chat model is a brilliant consultant. You can ask it for advice and it will give you a genuinely good answer, but you still have to go do the work. An agent is hiring that brilliant person in house, so they go do it too. He goes further and argues chat models cannot actually take meaningful work off your plate at all. A research task, maybe. Real work, no.
Chat still has a place, and both of them are clear about where. Mark uses it for hotel and food decisions, screenshotting options and asking which is better positioned for his day. Remy took a screenshot of Google Maps on the subway and asked Claude whether an Australian could pay with Apple Pay on the New York subway. Little questions to answers. As he puts it, using an agent to ask how many calories are in a banana is "killing a fly with a bazooka."
The illustration that lands is a marketing campaign, drawn from the agency work that used to be his bread and butter. Say you are running a Christmas in July campaign for an ecommerce brand.
The old way: the brief lives in Notion. You copy it, paste it into ChatGPT, ask for ads and email copy. You take that copy into Klaviyo or hand it to your email designer. You make the ads in some other tool. You paste the content into your content calendar. Then you go into Meta Ads Manager and manually enter all of it. Mark names the real cost: switching cost, plus a hundred small optimizations, spread across two, three, four, five different people. Remy adds the detail only someone who has done it would volunteer: uploading a hundred ads to Ads Manager took him an hour of pure data entry, most of it toggling off creative enhancements in a user interface he describes as awful.
The new way is one instruction. Tell Claude to kick off the Christmas in July campaign. It goes into Notion for the campaign outline, the discount, the offer. It goes into Shopify and pulls real product images. It has your brand guidelines, so the email comes back branded. It writes the ads, using an image generator like Higgsfield for the statics. It hands you five emails and a hundred ads. You pick your favorite email and your top ten ads, approve, and it goes into Klaviyo and Ads Manager and sets everything up, leaving it as a draft so you can hand optimize before clicking go.
Chat (stage one)
Agents (stage two)
Shape of the interaction
Question to answer
Goal to result
Who does the work
You do, after it answers
It does, then hands you the output
Analogy
A brilliant consultant you call
That same person, hired in house
Good for
Calories in a banana, is Apple Pay accepted, quick decisions
Anything that is real work
Campaign example
Copy the brief out, paste the copy back, tool to tool by hand
One instruction, Notion to Shopify to Klaviyo to Ads Manager
Memory of you
A black box file you cannot see or edit
Markdown files you own and carry between tools
The observe, think, act loop
Remy is careful to say he is not a software engineer and is not going to lose anyone in the internals. What he wants understood is the loop, because everything else in the episode hangs off it.
When you hand an agent a task, it runs three steps over and over until the task is done: observe, think, act. Or in plain language: look at the situation, decide the next step, do it. Look at the situation, decide the next step, do it. Then, when it judges the work complete, it outputs the result.
How does it decide it is done? Entirely from what you told it in the goal. His example: "research the top 10 business podcasts I should go on and put it in a 10 page PowerPoint deck." The agent will not stop looping until it has ten business podcasts in a PowerPoint. Be vague about what the finished thing looks like and the stopping point becomes the agent's judgment rather than yours, which is how you end up with an output you did not want. Defining what done looks like is the first discipline.
The application that runs this loop is called an agent harness. Claude Code, Claude's co-work mode, Codex, Manus, and Perplexity's computer product are all just different harnesses. This matters for reading the internet honestly: when someone online says "I built an agent that does this," taken literally that would mean they built a competitor to Claude Code. They did not. They tuned an agent.
And out of the box every one of those harnesses runs the same loop, which means there are exactly four levers you can pull to make it better, and each one attaches to a specific step:
Observe: context. The information you give it.
Observe: skills. The processes you have written down for it to follow.
Think: the model. The large language model doing the deciding, GPT‑5.5 or Opus 4.8 or whatever is current.
Act: tools. Out of the box a harness has a handful, usually including web search. If you want it handling your email inbox, you tune the act step by giving it more.
Mark reaches for a motorsport analogy and Remy improves it. Yes, the model is the engine, a Lamborghini versus a Volvo. But "the Volvo could beat the Lamborghini if the context, the skills, and the tools are really tuned and super super dialed." Then he lands the analogy he actually likes: today you are learning to drive. Manus, Perplexity, Claude, Codex are all just different cars. Codex is a Ferrari, Claude Code is a Range Rover, Manus is a G Wagon. Once you know steering, brakes, and accelerator, you can get into any of them and be productive in five minutes. The indicator might be on the other side, one might have cruise control the others lack, but the fundamentals transfer. Mark calls it the hot take of the episode: everyone is bullish on a specific model, and the skill that actually matters is model independent.
Figure 1. The loop every agent harness runs, and the four places you can change its behavior. Claude Code, Codex, and Manus all ship the same loop. What differs between a useless agent and a great one is entirely context, skills, model, and tools.
Live demo: the loop, running
Rather than describe the loop further, Remy runs one on screen in Claude Code. The prompt: give me a report of my best performing Instagram videos with a full breakdown of the angle, the hook, the topic, the title hooks, and why the video worked, using my Instagram handle, and create an HTML report.
He narrates the loop as it happens, and the play by play is the clearest part of the episode:
Act: it searches his computer for information about title hooks. Observe: it found files that look exactly right.
Think, act: read those two files. Observe: "this is a gold mine, there's a complete data set of 60 reels matched with play counts." He notes he had probably downloaded his reel performance for another project and the agent stumbled onto it. Without that, the loop would simply have continued and called a tool to go scrape Instagram instead.
Observe: all 60 reels matched with play counts, engagement rates, dates, pattern tags. But "the one field that genuinely isn't in this data is the spoken hook, first 3 to 5 seconds." Mark: "that's the bottleneck."
Think, act: it launches a subagent, which Remy defuses immediately as far less sci fi than it sounds. It is just another copy of itself sent to go check one thing.
Observe: "the data is rich and complete, the spoken hook retrieval is running, let me build the full report." Act: build the report.
The output is a styled HTML report titled "what actually works on AI with Remy," analyzing all 60 reels with the top one over a million views, categorized, each with its title hook broken out, plus a full ranked table by plays with dates, titles, hooks, and durations. Mark immediately mines it for content strategy, pointing at a "right versus wrong way to use Claude" entry and noting the whole report is a list of proven angles you can re run as new videos. Remy adds his own use case: if you do brand deals, generate exactly this report per brand so the client gets a performance breakdown of the video you made for them.
His estimate is that the run took roughly 20 to 30 loops and about 20 minutes of wall clock time. That is the honest shape of agent work: you give it a goal, it grinds, you come back.
You are not building an agent. You are onboarding an employee.
This is the frame that carries the rest of the episode. Open any harness raw and give it a task and it will be capable but will not do things the way you want, and the output will disappoint you. That gap is exactly why people try these tools once and quit. The fix is to stop treating setup like prompting and start treating it like onboarding.
Think of the agent as an extremely capable stranger. Everyone has that friend who is genuinely smart, good at solving problems, and will keep working until the thing is done. You now get to hire that person for about $100 a month, and he has infinite bandwidth. But no matter how capable he is:
unless he fully understands you and your business, or your role if you are an employee, he is useless;
unless he has the tools to do the work, an email account, a browser, a computer, he is useless;
unless he is trained on your secret sauce and your processes, he is useless.
Which maps exactly onto the three things the rest of the episode builds: context, tools, skills. Context is information: who you are, the business, brand voice. Tools are a computer, a browser, Gmail, Notion, Slack. Skills are your SOPs, the documents you would hand a virtual assistant. Get those three right and, in Remy's words, you have the best employee you ever had.
Context, part one: markdown files you own
Context is everything the agent should already know before you give it a task. Remy's own set looks like this:
About me. Everything about him as a person, including his story and his come up, so that when it writes landing page copy or a sales letter script it can reference real events from his life.
Business information.
Offer catalog, with every service and offer, the prices, and how they link together as upsells and downsells.
Ideal Customer profile, which he calls one of his most important files.
Brand voice, values, and how I like to work.
The comparison he keeps returning to: onboarding a real employee means handing them documents. You are doing literally the same thing to the agent. The only difference is format, and the format is markdown.
For anyone who has not met it, his definition is exact and unpretentious: markdown is one step above a .txt file. A text file is plain text with no headings, no bold, nothing. Markdown is a plain text file where you also get headings, bold, italics, and bullet points. If you use Notion or ClickUp, their docs are markdown underneath. Nothing more exotic than that.
Why it matters is mechanical rather than aesthetic. You could hand the agent a .docx or a PDF, but the agent then has to unpack the file and extract the text, because of how those formats are built. Markdown is a straight ingest. Easy for the agent means faster for you and cheaper in model cost. Mark asks how many markdown files Remy has on his computer. The answer is "like hundreds."
Is more granular better? Not necessarily. He has no theory here, just a preference for separate files per area, and he could have put them all in one doc. What matters more is quality, because these files are assets for your company. The warning he gives is specific and worth quoting in spirit: if you rush the ideal Customer profile and let a chat model produce slop you never read, one line might say your Customer's problem is red skin. If you sell a skincare product that does not address redness, you have now silently pointed every piece of marketing copy your agent writes at the wrong angle. Mark's rule of thumb: minimize subjectivity in the markdown files, keep them objective, keep them the kind of thing you would genuinely use as an input to a decision. Remy has Claude build all his files, and he reads every line before approving them.
The build method is the best practical tip in the section. He opened a new chat and said, roughly: I want you to build out a bunch of markdown files for me, an about me, an ideal Customer profile, and this list of others, and to do it, interview me to extract the information from my brain. Then he sat there for three hours answering questions using Wispr Flow, the voice to text tool, which cleaned up his rambling into formatted text. Three hours of talking, and the output was a set of files that are now company assets.
Why ChatGPT's memory is a trap
Remy makes a sharp distinction here that reframes the whole "but ChatGPT knows me so well" objection. People say they cannot move to Claude because ChatGPT knows them. He agrees that it does, and then explains what is actually happening: in the background there is a file, effectively a memory.md or about_user.md, that the product is writing on your behalf. You cannot see it, you cannot edit it, and you cannot take it with you.
Two problems follow. First, you do not control what gets saved. Tell it you just moved to New York while asking about cafes and that fact gets written down. Ask for advice about breaking up with your girlfriend and then ask for landing page copy, and the model may pull threads from one into the other. Second, and worse for anyone building a system, you cannot leave. The context is trapped in one vendor's product.
Owning your context in markdown solves both. Remy's folder goes with him. He opens it in Claude Code, Codex, or Manus and every one of them immediately knows everything about him. He stays mobile.
Building the folder live: empty, v1, v2
He demonstrates this from zero, which is why it lands.
He creates an empty folder called demo on his desktop and opens it in Claude Code, which scopes the agent to work in that folder the way a ChatGPT project scopes a chat. He has been using Claude Code for eight months. He asks it: who am I? It does not know. It tries to read a memory.md, finds nothing, and comes back with only his name and email, which it got from the computer's user account. Eight months of use, zero knowledge, because the context lives in files and this folder has none. As he puts it, it is not a black box.
Version one of the setup is one file. The most important markdown file when dealing with agents is AGENTS.md, though as Remy notes with visible amusement, "Claude have just decided to be difficult and in their systems they call it a CLAUDE.md." Most other harnesses, including Codex, use AGENTS.md. The behavior is the same: any file with that name in the folder gets loaded automatically at the start of every new session.
This is where he explains sessions and context rot. A session is a fresh start, the equivalent of a new chat. Each one has a fixed amount of room, which you can picture as the agent's working brain. As a session fills up, you get context rot, and the symptom is familiar to anyone who has kept one ChatGPT thread alive for hours: it starts hallucinating and forgetting things. What is happening is that the stuff you told it at the beginning falls out the top. Mark's response is one of the best lines in the episode: "that was V1 Mark, bro, I'm not on V1 Mark anymore," followed by his estimate that 99.999% of the world is still living in one endless chat.
So he drops a pre prepared CLAUDE.md into the folder and opens it on camera. Its contents: Remy's AI operating system. Who am I. The business. The brand voice. The ideal Customer. The offer. The values. How I like to work, preferences. He starts a new session in the same folder and asks who am I again, and this time it knows exactly who he is, what he does, his background, the brand, the three products.
Version two adds a context folder. The other markdown files do not autoload, so the CLAUDE.md has to point at them. He pastes in a context/ folder containing my-values, my-offer, ideal-customer-profile, brand-voice, about-me, and memory.md. Then he opens the CLAUDE.md and shows the pointer, and the wording is worth copying verbatim: "My context, read before any task. Everything you should know about me and the business lives in the context folder. Load the relevant files in before starting any tasks. Assumptions are the enemy. If the answer isn't there, ask me."
He also explains why the CLAUDE.md is not optional even though the context folder exists. Delete the CLAUDE.md and leave the folder, and the agent will not know you immediately. It will burn a loop wondering, another loop listing folders, another reading the context folder, and only then load it. Slow, and unnecessary. The autoloaded file is the shortcut.
The modularity point follows naturally. All of this is swappable. People who use Obsidian as a second brain are doing exactly this, just with a vault instead of a context folder. If you wanted that, you would delete the context folder and point your CLAUDE.md at the Obsidian vault instead. It all stems from one file.
The test drive: he asks the agent to write a cold email to Mark from Open Residency pitching himself as a guest, dictating the prompt with Wispr Flow. It runs a couple of loops, researches who Mark is, and comes back with subject lines like "operating a 20 person team without hiring 20 people" and body copy referencing "the operator as last line of defense" framing from the show, plus his own positioning about teaching founders to stop drowning in tools. Remy's verdict on his own demo is refreshingly blunt: "terrible cold email, by the way. Terrible." The point was never the copy. The point was that with zero additional instruction, it wrote in his philosophy and about his offers, because the context folder was rich. Mark's counter, delivered deadpan: "you're going to need to dial that email in a little better for me to say yes."
Context, part two: a memory system that improves itself
Off the rip the agent knows everything in the files. But what happens when you correct it? That correction is also context, and it needs somewhere to live.
Remy's memory.md is a plain list. Its header says: memory, learned rules, lessons Remy has taught me, I check these before every task. Underneath sit rules like "never write formally" and "always use real examples, never generic placeholder ones." The CLAUDE.md points at it with an instruction that is the whole mechanism in three clauses: your memory lives in this file; when I correct you, write lessons into this memory file as a new rule; check it before every task. In his words, "this is how you go from good assistant to knows exactly how I like things done."
He tests it live. He types: "please in future when creating websites I want you to always use dark mode and never ever use light mode." The agent replies that it will add this to the memory rules so it never forgets, reads memory.md, edits it, and appends the line. He opens the file in the folder and there it is: when building any websites, always use dark mode, never use light mode.
Two details he adds are worth keeping:
The @ symbol is a shortcut worth learning. Tag a file with @ and it behaves like a symlink, loading straight in. Reference it in prose without the tag and the agent has to burn a loop finding and reading it. Mark's phrasing, which Remy confirms: it zips right into the file.
Why build this when harnesses now ship memory. They did not until recently, so he built it himself. Now that Claude has memory built in, he still uses his own, because Claude's version saves to Claude's desktop app. If you want to open the same folder in Codex, you do not own that memory. His system is portable, theirs is not.
The obvious objection is growth. What happens when the file has hundreds of contradictory lines? His answer is unbothered: in six months, open a session and say audit my memories, take out anything contradictory or conflicting, clean it up.
He closes the section with a measurement most people never take. Claude Code has a built in /context command that renders a bar chart of the current session's context window. He runs it in the demo folder: 6% used, on a session where the only message sent was one short instruction. That 6% is the CLAUDE.md plus the pointed at files, loaded before any work starts. The lesson is a budget, not a warning: if you have a hundred bloated markdown files, you might start every session already 20% full. Keep the files lean and dense with high value information.
His own recap of context 101: the CLAUDE.md that Claude autoloads every session, an extra context folder that the CLAUDE.md points at, and a memory file that turns corrections into permanent rules.
Figure 2. The folder Remy builds live on camera, at each stage. The single autoloaded file is what turns a capable stranger into someone who knows you; the context folder is where the detail lives; the memory file is how corrections survive the session.
MCP: the universal translator
Context makes the stranger informed. It does not make him useful, because he still has no email account, no browser, and no phone. Tools are the act step, and the way you attach them is MCP, the Model Context Protocol.
Remy is emphatic that the name sounds far more technical than the thing is. All MCP does is give your agent a way to connect to third party tools: Slack, Gmail, Calendar, whatever.
The explanation he uses is one he explicitly credits to someone else, Ross Mike, whose framing is the one that stuck in his head. Claude speaks English. Every tool speaks a different language: Gmail speaks French, Notion speaks Spanish, Slack speaks Chinese. Before MCP, connecting them meant custom development, because Claude effectively had to learn French. Mark supplies the adjectives: heavy, slow, expensive. MCP is a translator sitting in the middle, so Claude never learns French; it makes calls and they get relayed. It is an oversimplification, and as Remy says, nobody really needs to know what happens underneath. If you know what an API is, it is the same idea: a way for tools to talk to each other.
Then he shows how simple it is in practice, because the abstraction matters less than the two clicks. Open the Claude desktop app, go to customize, and there is connectors. Browse and nearly every major tool is there. He connects ClickUp live, which he notes both of them are power users of: click plus, one click login, select workspace, connect, done. Back in the desktop app, ClickUp is connected, with a full list of available actions like create tasks and create lists, and Claude can now pull from and update it while planning campaigns. Every action defaults to needs approval. As trust builds you can switch to always allow, which Remy has done for essentially everything. He points out the thing most people miss: that plus button and login flow is MCP. There is an MCP server behind it. You never touch anything technical.
The strategic consequence is the sharpest business idea in the episode. Remy now makes purchasing decisions on MCP availability. Choosing email marketing software from five options, any of them without an MCP is out of the question. He does not use front ends anymore. Google Calendar, Gmail, Slack, Stripe for payments, Cal.com for bookings, and he cannot remember the last time he opened any of those apps. They are all connected into Claude, which sits as an intelligence layer on top of the whole stack, and that layer is the only thing he interacts with.
He extends it into product advice. A friend was building a travel planning AI agent as a website you log into. Remy's counter: build it as an MCP instead. Then anyone using Claude adds you as a connector, and your purpose built MCP makes Claude better at planning trips by giving it the extra tools it needs, like easier access to flight search. Mark agrees and sharpens it: rather than offering a hundred things, be genuinely dialed on ten or twenty, and it becomes an external data play in conjunction with your vertical.
Both of them then say the quiet part. Mid market B2B SaaS is getting decimated. Mark says the things he was doing in those tools he now does inside Claude. Remy adds the structural problem: a lot of that software is priced on seats, headcount is being reduced, agents are doing the work, and you cannot bill an agent for a seat.
Security and permissions
Mark raises the obvious objection: people are nervous about connecting all of this. Remy's answer is a trade off calculation rather than a policy.
Yes, there is risk. An email might go to the wrong person without his approval. He has weighed that against the productivity gain and decided it is worth it, and he thinks everyone has to run that calculation for themselves. His analogy: you hear horror stories about agents doing things, and you hear the same stories about people getting hacked online, and nobody responds to that by not using the internet.
The practical guidance is graduated, and it is the part worth acting on:
Start with read only access. Give it the tools but not the ability to act. Let it read emails without being able to send. Increase permissions as trust grows.
Or let it work to the finish line without clicking the final button. Draft only. Mark's line is that data is one thing, but anything holding your credit card is the last line of defense.
Know that the harness matters. Remy notes that Claude Code and Codex, built by Anthropic and OpenAI, are very good at asking permission, sometimes annoyingly so when he is answering emails. Frameworks like Hermes and Open Claw are built to ask less.
Remy himself gives Claude full access to everything including Stripe, has done for eight months of power use, and has never had an issue. Mark flags 1Password as a bridge too far for most people, and Remy's counterpoint is a genuinely reassuring detail: when the agent needs a password on his machine, macOS pops Touch ID and he has to approve it with his finger. Mark, converted: "I stand corrected." Remy: "I feel like Iron Man or something."
The second half of the security question is the one enterprises actually get stuck on, and Remy is candid that it is unsolved for him. Some large companies he works with have said flatly that no personal information or company IP may run through a Claude model, not because of agent actions but because they are not convinced about what happens to their data. He separates that cleanly from the personal case and moves on.
The starter tool stack: six connections
Back to the loop. Context tuned the observe step; these tune the act step. Remy's list of what everyone should have connected, in the order he gives them:
1. Apify. A marketplace of thousands of scrapers other people have built. He scrolls it on camera: Instagram profile scraper, YouTube transcript scraper, Reddit post scraper, LinkedIn post scraper, Google Maps scraper, Craigslist scrapers, lead scrapers, email scrapers. Mark reacts to the implication immediately, noting people pick a model based on its ability to scrape one social network when this covers everything. Remy's claim: "there's almost nothing that Claude can't scrape off the internet for you with Apify connected." The free plan gives roughly $5 a month of usage. This is the one he says everyone needs regardless of which harness they use, and later, in the quickfire round, the one he names as the single connection that changed the game.
2. Firecrawl. Every harness ships with a web search tool, but web search is primitive. It finds facts. Ask for the price of Bitcoin and it works. Ask it to look at a website and audit the branding, and it cannot, because web search cannot read the page. Firecrawl can. It scrapes the whole page, sees the buttons and where they are, sees the brand colors. Remy's example of the difference: without it, "here's my ecommerce store, run a conversion rate optimization audit" gets you a search about what the company sells. With it, you get feedback like your buttons are the same color as the background, make them brighter. Mark, who thinks in merchandising, jumps on this as the highest value one for ecommerce.
3. Composio. This is the one he introduces as the real find, with the backstory that produced it. He had all his connectors built into Claude. Claude went down for an hour one week, he had work to do, and he opened Codex instead. His folder came across fine, because he owns his context and skills rather than embedding them in Claude. What did not come across were roughly 50 connected tools. Composio is the fix: thousands of apps, one click login, and then a single MCP that carries all your other MCPs underneath it. Mark's summary, which Remy accepts: "it's an MCP for MCPs." It has a second advantage he calls out specifically: Claude's desktop app lets you connect one Gmail account, and most people have several. In Composio he has three active Gmail connections. It does not appear in Claude's browse list, so you add it as a custom connector with a URL. The payoff is portability. If he wants to try a new harness tomorrow, he connects Composio once and everything is there.
4. Chrome DevTools MCP. Built by Google Chrome. It lets Claude spin up its own browser, which matters most for checking its own work. Ask it to build a website to your brand guidelines and the first attempt may be wrong in ways only visible on screen. With a browser, it can look at what it built and decide to keep working. His other use: a founder registry he saw on X had a twenty minute application form, so he pasted the link into Claude and asked it to apply on his behalf while he worked on something else. Because it had his context, it answered every question in his voice.
5. Playwright MCP. Same idea, different trade off. Chrome DevTools spins up a fresh browser with none of your logins, so "send 100 LinkedIn connection requests to this list" ends at a login screen, and a login you do enter is gone next week. Playwright drives your current browser instead, with your sessions intact.
6. Higgsfield. Claude has no image model and no video generation, and neither does ChatGPT for video. Higgsfield is, in his framing, the Composio of image models: it pulls in Google's top models, OpenAI's, and the Chinese models that have gotten very good, and exposes them through one MCP. He credits the company's marketing as something that "needs to be studied," saying they have made themselves the default name for AI content through influencer marketing.
The demo for Higgsfield is a full ad pipeline in one prompt. Using a creatine gummies brand as an example, he gave it a product image and said: scrape competitors' ads running in the Meta Ads Library and generate 20 static image ads using the Higgsfield MCP. It used Firecrawl to scrape the ads library, generated the statics, and it is also connected to Meta through an official connector Meta has released for agents. He walks the output candidly, and the honesty is the useful part: with ads you generate a hundred and five to ten are good. One product looks squished. One he would run as is. One he calls phenomenal. The gummies do not look perfect, and he notes the fix is simply to supply a real photo of the gummies. All of this off a basic prompt; a serious version would have skills built around static ads, headline writing, and copywriting.
The mental model that ties the six together is the same employee metaphor. Your capable hire sits at a desk. Next to him is a bookshelf of SOPs, which are skills. On his screen are his applications, which are these tools. You never say "go use Apify." You say scrape this brand's Instagram page and give me the best performing videos, and he looks at what he has and picks the right one. Remy also notes what happens without them: the Instagram demo at the start of the episode used Apify under the hood, and without it the agent would have come back asking him to send the videos manually.
Image and video models from several vendors through one connection
Claude generates no images; nothing generates video natively
There is a meta lesson buried in this section that both of them stop to underline, and it may be the most useful thing in the episode for a beginner. Mark admits his instinct would be that switching to a new harness is hard, then realizes he could simply ask ChatGPT how to connect Composio. Remy's response: "that's how I learned all this. All this stuff I'm sharing with you today, it's pretty much all learned from asking Claude." His follow up is the mechanical version: if you follow the steps and get stuck, screenshot the error, paste it in, and say what do I do. Mark calls that a cheat code on its own. The behavioral prescription he adds: if you are not technical and would normally ask a teammate to do something small, force yourself not to. It might take you ten minutes instead of their two. Figuring it out yourself is the unlock.
Skills: SOPs for your AI
Remy calls skills probably the most important thing to master. The stranger now knows you and has tools. What he does not have is your processes.
The definition is deliberately unglamorous: a skill is an SOP for AI. If you had a virtual assistant doing Customer support, you would write a Google doc titled Customer support triage workflow, and it would say step one, log into monday.com; step two, navigate to this section; step three, find tickets. Foolproof, so they cannot get it wrong. A skill is exactly that document, written for an agent, in markdown instead of Google Docs.
The problem it solves is repetition. Ask an agent for a client proposal with the price and offer, and it will hand back a PDF that is very rarely right first time. You go back and forth: the price should be at the bottom, not the top; move the logo to the top left; change the text from blue to black. After fifteen or twenty minutes you have a proposal you are happy with. Then next week you need another proposal, you start a fresh session, and every one of those preferences is gone. You could push them into memory.md, but as Remy says, that is the wrong place; you would be building a de facto skill. Build the real thing instead. Tell Claude to save the process you just went through as a skill, and it packages up the steps, your corrections, and often the finished PDF as a reference. From then on it follows that SOP to a tee.
Mark's example of the simplest possible useful skill: social media captions. Hook, body, call to action, tone of voice, not too professional, no hashtags, and ten example outputs. Nail it once, and you are done.
The anatomy of a skill
He opens a real one. Inside his OS folder there is a hidden .claude folder, which he immediately defuses as nothing more than Claude's settings for that project. Inside it, skills/. Inside that, one folder per skill. He opens YouTube titles.
A skill is a folder containing a skill.md, and every skill has three parts:
Name. "YouTube titles."
Description. "Generate YouTube title options for a video topic. Use when the user says generate titles, YouTube titles, or title options for my video."
Contents. Everything below the two lines that separate the header from the body.
The bookshelf metaphor is what makes this stick. Your agent has a bookshelf next to his desk. On the spine of every book is just the name and the description. When you ask for YouTube titles, he scans the spines, sees one whose description says use when the user says titles, pulls it out, reads it, and loads the contents in.
This is progressive disclosure, and it is the reason the system scales. Only the names and descriptions of every skill load at session start, which is why a hundred skills do not fill your context window. If the full contents of every skill loaded, the bar would be full before you typed a word. The contents load only when the task calls for them.
Figure 3. Progressive disclosure, the mechanism that lets a single agent carry hundreds of processes. Only the spines load at session start. The body of a skill loads when the task reaches for it.
Inside the YouTube titles skill, the contents read exactly like an SOP for a person. Step one, get the topic: if the user gives it to you, use it, otherwise extract it from the folder or file name. Step two, competitor research: use Apify to scrape top performing videos on the same topic, with target channels he specified, extracting the video titles, title length, structure, use of numbers, and capitalization patterns, plus instructions on how to use Apify. Step four, generate the title options across categories: curiosity gap, benefit first, contrarian, number.
Two asides matter here. He did not write this skill; Claude did, and Claude will always build skills for you. And this is where your IP actually lives. Ask a fresh Claude session to write YouTube titles and the output is generic and bad, and Remy is confident enough to say "I know you've tried it." He has a way he thinks about titles that is specific to him, and putting that into a skill is packaging his judgment. Same for a media buyer: ask Claude to handle your Facebook ads and it does not know your unique sauce. Mark takes the point further, and it is the strongest business claim in the episode: enterprise value is going to live in what your skills and markdown files are. Asking someone what theirs look like is a good probe for how deep they actually are.
Two ways to build a skill
Goal first. Say up front: build me a brand guidelines skill, here is the brand guidelines PDF. It goes and builds it. Claude and essentially every other harness ships with a skill creator skill, a skill whose contents are a step by step guide for writing more skills, and you do not have to download it. So when you ask, the agent looks at its bookshelf, finds the skill creator, reads it, and knows what to do.
Process first, which is how he builds 90% of his skills, and which Mark reframes as bottom up rather than top down. You do the work once, conversationally, and then freeze it. His example is the ad scraper he built:
Fresh session: here is the Meta Ads Library link for this competitor, use Firecrawl to scrape their entire ad library and download it. Send.
Now analyze every one of these ads and break down the metrics I care about. Send.
Now find all the different landing pages they are running ads to. It comes back with five.
Use Firecrawl to screenshot all the landing pages and build a master doc breaking down each one.
Having taken Claude on the journey and gotten a result he liked, he said: create a skill for this process we just did and call it ad scraper. It used the skill creator and packaged the whole thing.
The ads analyst skill, running
He demonstrates it live. Type / in Claude Code and every available skill appears; he picks ads analyst, described as "orchestrate competitive ad research, extract ads, create a report." The only input is an ads library URL. He runs it against a friend and competitor in the AI space, and because it takes about twenty minutes, he also shows a run he did earlier with the same URL.
The output landed in an active/ folder inside his OS folder, which is where he keeps miscellaneous work that does not belong to a pillar. Inside, the skill had produced:
a master report in markdown, which he notes looks exactly like a doc you would see in ClickUp or Notion;
a linked version with links out to the individual ads;
every asset downloaded, including essentially every video ad;
a full landing page analysis;
a final report in HTML.
The HTML report is the payoff: 54 active ads, 37 video and 14 image, the price point, the longest running ad, the newest funnel from three days ago. Then a section called things to steal or beat. Two of its findings are good enough that both of them stop on them. Reframe the category to justify the price: the competitor sells what is technically a Notion template but positions it as an operating system. Name your mechanism: calling the method context engineering makes it feel ownable.
Then he makes the point about iteration that keeps this honest. Looking at his own output, he says he does not like how the landing pages are formatted, too big and full screen. The fix is not to redo the work. It is to tell Claude he wants them smaller and better fitting and to update the skill. Next run, that problem is gone. The V1 of a skill is never perfect. Most of the skills he uses are on V5 to V10. You keep sanding until it is right.
The product shoot skill
A thirty second detour that he introduces as blowing people's minds, and it earns it. He built a product shoot skill for a friend's furniture brand. The input is a reference image of the product, which can be nothing more than an iPhone photo, in this case a cabinet. Run the skill pointed at the image and it generates the full set of production ready images: the same cabinet rendered as an editorial shot in a styled office, a different angle with someone's hand touching it, a close up of the veining on the marble. The generation runs through the Higgsfield MCP, using what he believes was Google's Nano Banana Pro as the image model. He calls it one of his favorite skills he has ever built.
Building a skill live, with beehiiv
Then the main event, built from nothing on camera. The goal: a skill that writes subject lines for his newsletter by looking at what has historically performed best in beehiiv.
Step one, connect the tool. beehiiv is not a native Claude connector, so he uses add custom connector, names it beehiiv, and needs a server URL. He goes into his beehiiv account, into MCP settings, and copies the URL they publish. His note for anyone doing this with another tool: most software that ships an MCP publishes a URL exactly like this, and if you cannot find it, ask Claude to walk you through it. Mark adds that beehiiv was one of the first email platforms to ship an MCP connector.
Step two, do the work conversationally. In a fresh session in his OS folder he asks, in plain speech: go into Notion and find my most recent newsletter, which is not posted yet, then use the beehiiv MCP to find all my subject lines and pull in all the data so you can see which perform best, and use that to write ten new options for the most recent newsletter. He points out that his prompts have no structure or formatting at all. It is just natural language dictated through Wispr Flow, and he thinks the anxiety about prompting correctly is fading fast.
The output does what you would want and a bit more. It extracted the patterns that win from his own history: do not stack tool names, because his two worst subject lines both jammed two tools into one line; "I plus bold stance" wins, "I plus boring task" does not. It cites winners like "I don't want my business to be trapped in Claude," and flags an honest miss, "I automated Meta ads in 20 minutes," which he thinks is a good line that did not perform. It even picked up that he opens every subject line with a particular emoji.
Mark spends a moment on the wider possibility, which is a good instinct: with that MCP you could run cohort analysis, extract lists of operators or athletes from your list, follow UTM tracking down to which individual subscriber clicked which email, and attribute subscribers back to the lead magnet that produced them. Remy confirms he uses it constantly, including one occasion where he migrated his entire beehiiv list into Resend with a single instruction.
Step three, freeze it. He asks Claude to build a subject lines skill that always goes into beehiiv to check the latest and historic best performing subject lines and uses that data to write for the newest newsletter. Claude writes the skill.md with a name, a description destined for the spine, and a body: step one, load the tools; step two, pull beehiiv performance data; step three, find the draft and the hero story in Notion; then write ten ranked options and use this output structure. As Remy says, it is so simple.
Step four, put it where it belongs. He asks Claude to save the skill into the demo folder from earlier, and the folder gains a .claude/ directory containing skills/subject-lines/skill.md. Mark asks why the folder looks dark, which gives Remy the chance to explain that a leading dot means hidden, that Command Shift Period toggles hidden files on a Mac, and that he deliberately keeps his settings and skills inside .claude/ so his working folder shows only the files he actually cares about.
Orchestrator skills
The last piece is chaining. He has a skill called YouTube publish workflow whose entire skill.md says: this is an orchestrator skill, use these other four skills in sequence. Step one, use the YouTube title skill. Step two, use the YouTube thumbnail skill. And so on. The skills it calls are not nested inside it; they are standalone skills sitting on the same shelf. The ads analyst he demoed earlier is itself an orchestrator, calling an ad creative analysis skill and a creative strategy skill among others.
Mark's read on this is the answer to everyone worried about being made irrelevant: the strategic mind still decides the sequencing. Some people want thumbnails before titles, and without you specifying that, the agent has no way to know. The chain is a piece of judgment you encode.
The Agent OS folder structure
Mark asks for the zoom out: what does the machine look like after you have put a lot of hours in. Remy gives a live walkthrough of his actual setup, and opens with what he stopped doing.
He used to build granular, niche agents, and he is careful to explain that "agent" here just meant a folder. He had a marketing folder, an ads folder, maybe even a subject lines folder, and he would open Claude Code scoped to whichever one matched the work. The more he built, the more annoying and inefficient that became. Now he runs everything out of one OS folder. OS stands for operating system.
He is blunt about why people do the other thing. As humans we love to treat these as people, giving them names and roles, and it gets clicks on Instagram: "my team of agents did this." In reality they are folders. More folders and more specialization means more moving parts to worry about, and the models are now smart enough that the specialization buys you nothing. He runs everything from the OS folder and puts the effort into building really good skills instead.
The structure, top to bottom:
OS/ at the top, the holding company.
Inside it, one folder per pillar: a business, a brand, a major area of life. For Remy that is his education and consulting business plus the client work he does. For Mark it would be Iconic and Open Residency, plus probably a personal/ folder for health.
Plus an active/ folder for one off projects and experiments that do not fit a pillar. He created a folder there just to prep for this podcast. If something loose turns out to matter, it graduates into a pillar.
Inside each pillar, the working folders: content, finance, legal, newsletter, website, marketing.
The navigation is where the muscle memory lives, and both of them slow down for it. Across harnesses the conventions are the same: slash brings up skills and commands, including built in settings like /model for switching the model behind the loop, and at brings up files and folders to tag. He works with the whole OS folder open and tags what is relevant: type @ and start typing marketing, and his marketing folder appears. Select it, then give the task, "write me 10 new ads for my upcoming workshop," and the agent works in that folder without him ever having scoped the session down to it. Mark's fifth grade translation, which Remy endorses: you are loading the macro folder so it gets everything, then telling it where to look first.
Does the agent pick up skills automatically, or must you invoke them? Both, depending on the skill. Ask for a hook for an Instagram video and it finds the hook skill on its own. But the ads analyst takes a bare ads library URL as input, so dropping the link in with no instruction leaves it wondering what on earth to do with this. Naming conventions matter, and so does knowing your own structure. Mark's line: it is like having 50 employees, you should know their names.
Figure 4. The holding company model. One operating system folder at the top, pillars underneath, and context, tools, and skills that live at whichever level they actually apply to.
The global versus project distinction is the last structural idea, and the holding company analogy is exact. Global is the holding company: things that apply to every company underneath. Project level is the subsidiary. His global CLAUDE.md in the OS folder holds preferences about himself that will never change no matter what he works on, including that he is a non technical founder who is learning, plus months of accumulated lessons and his core principles. Mark spots one on screen and Remy reads it out: "boil the ocean, the marginal cost of completeness is near zero." It tells the agent that the standard is not good enough, the standard is holy hell that is done.
Then the project level. The podcast folder he built to prep for this episode has its own CLAUDE.md, and it says almost nothing: this workspace is Remy's base for podcast appearances, and this folder is for prep, demos, diagrams, and post episode analysis. That is all it needs to say, because the global file carries the rest. Working from the OS folder and tagging @podcasts, both files stack. He notes the agent would find the folder even without the tag; he tags it for cleanliness and peace of mind. On screen, it produces his episode plan.
Skills split the same way. Global skills go anywhere. The product shoot skill he built for the furniture brand lives inside that brand's folder, because he will never use it anywhere else.
Mark asks how he would architect a health system, since he is building one: DEXA scans, blood work, macros, an Oura sleep score with an MCP, an Eight Sleep integration. Remy's answer takes fifteen seconds and involves no cleverness at all: OS folder, a personal/ pillar, a health/ folder inside it, dump everything in, DEXA scans and all, and write a CLAUDE.md for the health folder saying this is my health workspace where I analyze blood work, plus my health goals. That is literally it. Mark extends it to scheduling: with Gmail connected, the agent knows to email his preferred DEXA vendor every seven weeks to keep the cadence, and his lens on food has become saturated fat and sugar because that is what his blood work told him.
Remy's closing advice on structure is to not overthink it. Do not chase the perfect layout. Create a folder called OS, write your CLAUDE.md, add the context, and let the subfolders appear as you do real work. Do not let structure hold you up from starting.
What it actually costs
Mark asks the question most people are too polite to ask: what is the honest monthly bill? Remy's answer is about $1,000 a month, across a lot of tools, with the caveat that being on top of this is his job, he is constantly testing things, and he has subscriptions he forgot about months ago. Mark's immediate suggestion is that he needs a skill on a schedule to go through his email receipts and cancel what he does not use, which Remy concedes is a good idea. Mark reframes the number as roughly $12,000 a year for what functions like a large team.
Remy then makes an economic point worth flagging as his read rather than as fact. None of the big AI companies are profitable, or at least he believes none are; he explicitly says he does not know whether Anthropic has crossed over. His claim is that consumer plans are heavily subsidized against real compute cost, and his estimate is that a $300 a month Claude plan carries something like $6,000 of token usage. He sees the squeeze already: prices holding steady while usage allowances quietly shrink. He points at the newest model as evidence, saying it is not included in plans and has to be paid for at real cost, which he calls disgustingly high. Mark's take is that it is a great new revenue line and that anyone who wants the best model will pay.
His forecast: this cannot continue at a loss forever, prices will rise, and there will be a substantial market for open source models you run yourself, on hardware as modest as a couple of Mac minis. Mark's version of the same thought, extending the car metaphor he started the episode with: right now there are only Lamborghinis and Range Rovers, and the Kias and Hondas are coming.
Two layers: personal OS and business infrastructure
Everything so far has been layer one, the personal AI operating system. Remy describes it as an extension of you, a robot arm, and says he now completes his entire workday inside it. It handles his tasks and his grunt work.
Layer two is AI business infrastructure, and the reason it exists is that some agents do not belong on any individual's laptop. An agent handling Customer support should not live on the founder's machine, and probably should not live on a support rep's machine either. That is where more advanced frameworks like Hermes and Open Claw come in, and it is why, in his telling, Mac minis were flying off the shelves: people wanted a machine that was not theirs, a company computer for agents to run on.
Then he does the thing that makes this episode good. He remote accesses the Mac mini at his house, live, and opens the file system so you can see that the Hermes agent is also just a folder. Inside it: folders and markdown files, including a soul.md that tells it how to behave. Same principles, same three levers. When you graduate to running agents autonomously on another machine, nothing conceptual changes. Context, tools, skills. The frameworks are just better suited to running unattended.
Mark raises the question nobody has answered yet: who owns the files? If an employee builds a stock forecasting skill on their laptop and a colleague needs it, you can airdrop the file, but now there are two copies and they will drift apart within a week. Remy is honest that he has not found a clean solution. You can bootleg it with GitHub or Google Drive. Some startups are appearing around shared skill management and he needs to get his hands dirty there. And then the harder version of the question, which he explicitly declines to answer: skills are assets for the company, and employees are the right people to build them because they know the work, so does the employee walk away with them when they leave, or does the company own them? "It's a difficult one. I can't say."
Mark's instinct is a split that matches the global versus project structure: the CLAUDE.md layer should be macro and uniform across the company, the stable stuff that does not change, while skills run downstream into individual creativity, taste, and preference. Some people want it in copy, some are visual. Remy agrees, and repeats his standing advice for anyone starting: build your own skills first, cross the team management bridge later.
Making a company AI native, from the bottom up
The last substantial argument in the episode. Remy states the stakes first: every company needs to put AI at the forefront of all their operations or they are going to lose to a company that does. That line opens the episode as its cold open, and this is where he earns it.
There are two ways to do it, and he thinks one is more correct.
Top down means bringing in AI specialists to build for people. They interview each department, sit with employees, and try to build skills on their behalf or build agents to absorb roles. The flaw is depth of knowledge. The people doing the tasks know every ounce of nuance, and a specialist would have to understand all of it for every task in the company. That does not scale; you would need roughly one specialist per employee.
Bottom up means training your employees to be AI literate and getting them building their own personal operating systems. The goal is explicitly not to fire your team. Keep your best people, train them into 100x employees, then go on a hiring freeze and keep growing without adding headcount. Automation spreads from the inside, as each person automates their own job, until the company is saturated.
His suggestions for making it stick are cultural, not technical: build weeks, competitions, a skill of the month with a $1,000 cash prize, a weekly AI call where people share what they have built. Make it part of the culture.
Mark half agrees and pushes back usefully on two fronts. First, the incentive problem: some employees feel that the more they adopt, the more they are digging their own grave, so you have to talk openly about which skills actually do well in the new economy. Second, he thinks the right answer is a mix. At Iconic he used a six page document, about 60% of it uniform questions for everyone (what tools do you use, what processes exist, where are your bottlenecks) and 40% tailored to the individual employee. Just running that download gives you an instant pulse check on where everyone is, and for him it was "a big shot in the arm." His conclusion: pure bottom up is too slow, and a mix of both goes faster. Remy agrees completely.
Quickfire
The closing round, verbatim in substance:
First skill everyone should build? An email inbox triage skill, run on a schedule three times a day.
Biggest beginner mistake? Perfectionism and fear. People say they want to use Claude but are scared to even type a prompt in case they break something, because they see it as a black box.
One thing you automated this week most people would never think to? "My laundry."
Most overrated AI tool? The app builders, Lovable and Bolt and the rest. Just create a project folder inside your holding company called website or app and have Claude build it there.
Most underrated? Composio.
How long until a total beginner feels the productivity gain? A few days.
The task you will never let an agent touch? Thinking. And phone sales.
One MCP connection that changed the game? Apify, for the thousands of scrapers.
Finish the sentence: in two years, every founder will have a... Personal AI operating system.
Is Obsidian still relevant?Obsidian is a note taking app built on markdown, which he calls "the love language of AI." People use it to manage a second brain of hundreds or thousands of markdown files, effectively the context folder from earlier at scale. His verdict: good for solo operators, messy the moment you want a team to access the same files, and not needed for beginners.
One skill that changed the game? The Claude council skill, which he found from Ollie Leman on X and which became his most viral Instagram video. The premise is a real weakness of chat models: ask Claude why something is a good idea and it will build you a perfect case, then ask why it is a bad idea and it will build you an equally perfect case, which is why you cannot trust it with decisions. The skill spins up several Claude sessions with distinct personas, a contrarian hunting for what goes wrong, an optimist, five in total. It shuffles their answers anonymously and hands them to five fresh peer reviewing sessions. Then everything goes to one final session, the chairman, which reads all of it and gives the verdict. He uses it for decisions like whether to add a feature or what to work on next.
Favorite book? The Alchemist by Paulo Coelho, read four times. Favorite podcast: My First Million, and Open Residency.
A brand to give flowers to? Composio. A creator?Ask Catgpt, who inspired him to start creating content.
Where to start, in his words
Mark asks for the recap, and Remy's answer is a sequence you could run this afternoon:
Create the folder. Call it OS. That is the holding company, and it is the seed.
Build the context files. Open a fresh Claude or ChatGPT session and say: help me build markdown context files, an about me, an ideal Customer profile, product and services information, brand voice, and interview me to extract the answers from my brain. Answer the questions. Put the files in the folder.
Connect your tools. He says this could even be the first step. Tools and context together take a few hours.
Complete an entire workday from inside the harness. That is the first real goal, and it will be messy at the start.
Turn repeated work into skills. Any time you take a task off your plate and know you will need it again once or twice, tell the agent to make a skill out of it.
His framing for why this compounds: the system grows with you, and you can only grow it by working inside it.
Mark's challenge to the audience, delivered with the exact energy you would expect: "if you had made it this far and you don't take action, you are a loser." And his softer version: do even 10% of what was covered here and you are in a tiny fraction of people who have graduated past chat.
Key takeaways
Chat is question to answer. An agent is goal to result. Chat cannot take meaningful work off your plate, because you still have to go do the thing.
Every agent runs the same three step loop: observe, think, act, repeating until the goal's definition of done is met. It stops when your goal says to stop, so define what done looks like or the agent decides for you.
You do not build agents, you tune them. Claude Code, Codex, and Manus are all harnesses running that loop. The four levers are context, skills, the model, and tools, and three of the four are yours.
Own your context in markdown. ChatGPT's memory is a file you cannot see, edit, or take with you. A folder of markdown files opens in any harness and makes it know you instantly.
The one file that autoloads is the keystone.CLAUDE.md or AGENTS.md loads every session. It should point at everything else, including the memory file.
A memory file turns corrections into permanent rules. Tell it to write lessons there when you correct it, and to check it before every task.
Keep context lean. A near empty demo folder already spent 6% of the window before any work. A hundred bloated files and you start every session a fifth full.
MCP is a translator, not a technology you have to learn. One click in the connectors panel is MCP. Remy now refuses to buy software that lacks one.
Six connections cover most of it: Apify for scraping anything, Firecrawl for reading pages, Composio as one MCP holding all the others, Chrome DevTools and Playwright for browsers, Higgsfield for image and video.
Skills are SOPs, and progressive disclosure is why hundreds of them fit. Only names and descriptions load at session start; the body loads when the task reaches for it.
Build skills process first. Do the work once conversationally, then say make this a skill. Expect V1 to be wrong; his live skills are on V5 to V10.
Orchestrator skills call other skills in sequence, which is where your judgment about ordering gets encoded.
One OS folder, pillars underneath, holding company style. Global context and skills at the top, project specific ones inside the pillar, and they stack when you tag a folder.
Start read only. Give tools without permission to act, then escalate as trust builds.
Companies go AI native from the bottom up. Train the people who know the work to automate their own jobs, because no external specialist can learn every task's nuance.
Start today with a folder called OS, an interview to generate your context files, and your tools connected. Then try to complete one entire workday inside the harness.
An honest footnote
Two things are worth naming, neither of which undercuts the substance.
The episode is also a funnel. There is a free playbook PDF, a paid Claude Code plugin called AI Onboard that builds this whole structure for you at half price for the audience, an affiliate code for Apify, and read spots for Wispr Flow, Ketone-IQ, and Momentous. That is normal for the format, and the teaching is not withheld behind any of it; the entire system is described in enough detail to rebuild by hand. But when Remy says the underrated tool is Composio and the brand he gives flowers to is Composio, you are hearing genuine enthusiasm from someone whose business is being early on these tools, which is not the same thing as a controlled comparison.
The numbers are lived experience, not measurement. Five to ten times more productive, roughly $6,000 of token usage inside a $300 plan, and none of the labs being profitable are all offered as his read, and he says so about the profitability point explicitly. Treat them as directional. His security posture is also further out than most people should copy on day one: full write access to Stripe, 1Password connected, always allow across the board, justified by eight months without incident and a Touch ID prompt standing between the agent and a credential. His own advice to start read only is the better default, and it is the advice he gives rather than the practice he follows.
What holds up regardless of vendor is the architecture. Context, tools, and skills as three separate levers on one loop, files you own rather than memory you rent, and progressive disclosure as the reason a large library of processes stays cheap. Those ideas survive whichever harness wins.
Chapters
0:00:00 Trailer
0:01:21 The shift from chat to agents
0:04:04 What "agent" actually means
0:11:08 The observe-think-act loop
0:18:01 Live demo: analyzing an Instagram Reel
0:20:24 Tuning an agent vs building one
0:22:45 Context files and the CLAUDE.md
0:44:03 A memory system that improves itself
0:52:06 MCP connectors, the universal translator
0:59:21 Security and permissions
1:02:19 Starter tool stack, Apify to Composio
1:18:21 Skills as SOPs for your AI
1:47:34 The Agent OS folder structure
2:06:46 The future of teams and intrapreneurship
2:16:20 Quickfire questions
Notable quotes
"Every company needs to put AI at the forefront of all their operations or they're going to lose to a company that does."
Remy Gaskell, 0:00:05
"Chat is just question to answer, and an agent is just goal to result."
Remy Gaskell, 0:04:20
"Look at the situation, decide the next step, do it. Look at the situation, decide the next step, do it."
Remy Gaskell on the loop, 0:13:40
"The Volvo could beat the Lamborghini if the context, the skills, and the tools are really tuned and super super dialed."
Remy Gaskell, 0:16:50
"In order to make the most out of these agents and make them actually useful, you need to treat setting them up like onboarding a real employee."
Remy Gaskell, 0:21:00
"You have no control with these chat models about what it knows and what it saves."
Remy Gaskell on ChatGPT's memory, 0:33:00
"Load the relevant files in before starting any tasks. Assumptions are the enemy. If the answer isn't there, ask me."
from Remy's CLAUDE.md, read on screen, 0:38:00
"This is how you go from good assistant to knows exactly how I like things done."
Remy Gaskell on the memory file, 0:46:00
"Claude does not need to learn French and Spanish and Chinese. It just has this little translator in the middle."
Remy Gaskell on MCP, 0:55:00
"I don't use front ends anymore of apps. I complete a full workday from inside my Claude Code."
Remy Gaskell, 0:57:20
"Just start with read only access if you don't feel comfortable, and then you can escalate it up as you go."
Remy Gaskell, 1:00:30
"Enterprise value is going to live within what these skills and MD files are."
Mark Mastrandrea, 1:32:00
"The first time you build them, the V1 of the skill is never going to be perfect. Most of my skills that I use are on like V5 to 10."
Remy Gaskell, 1:35:00
"People go, I've got my team of agents doing this. And in reality, they're just folders."
Remy Gaskell, 1:52:00
"Boil the ocean. The marginal cost of completeness is near zero."
from Remy's global CLAUDE.md, 1:56:00
"The goal is not to fire your entire team. You want to keep your best people, train them so they can become like 100x employees."
Remy Gaskell, 2:10:30
"The task you'll never let an agent touch? Thinking. And phone sales."
Remy Gaskell, 2:16:40
"Obsidian is just a simple note-taking app that uses markdown, which is the love language of AI."
Remy Gaskell, 2:22:00
========================================
Every company needs to put AI at the
forefront of all their operations
[music] or they're going to lose to a
company that does.
>> That's Remy. He's a young non-technical
[music] founder who's a master at
explaining AI in a way you can actually
understand [music] and implement.
>> In order to make the most out of these
agents and make them actually useful,
you need to treat setting [music] them
up like onboarding a real employee. And
you do this through context, tools, and
skills. He shares his screen and shows
us exactly how [music] he builds AI
systems, timeless principles you can
steal, and the easiest way to get
started today.
>> If you're starting from zero, this is
all you need. Just this episode, [music]
and I think you're set.
Guys, in this episode, Remy is going to
demystify everything that's going on
with AI, the terms, the tools, the
models, everything. an a toz guide to
know what's going on today in AI. He has
spent the last two years living in this
stuff and he is about to embed in your
brain all things AI that he has learned
in the next couple hours. He's not some
technical engineer either. For the last
couple years, he's been running an ad
agency. This is all self-taught. And now
he's running the AI agent team for one
of the fastest growing ecom brands in
the world today. And I know this AI
stuff is very overwhelming. So, I want
you all to take a deep breath and by the
end of this episode, you're going to
know exactly what the next steps are.
Let's get into it. So,
>> I guess let's just start with, you know,
what are the kind of the two stages of
using AI.
>> So, stage one was chat, right? You'd be
pretty hardressed to find anyone,
especially founders and business owners
that aren't using chat models. So, think
like chat GBT or Claude Chat or Gemini.
But stage two is now agents. So we've
moved from chat to agents and the
founders and employees that are now
using agents are like no word of a lie
like five to 10 times more productive in
their day.
>> Probably more probably more than 5 to 10
to be honest with you.
>> Probably. I mean even if you take like
the the middle mark which is like seven
times more productive. Think about that.
That's a week in a working day and like
5 to seven weeks in a week. And then if
you stack that up over like a month a
year like it's toast. So, you're
starting to just get this massive divide
opening at the moment with the people
who are using agents and the people who
are stuck using chat and you want to be
on the agent side.
>> Yeah. I mean, it's unfortunate for some
of the people listening out here that
may or may not be employees, but like a
big thing we talk about is just opex,
your actual labor as a percentage of
revenue. I think everybody is racing to
just get this down the percentage of
revenue. And I feel like people are
doing it with just chat and they don't
understand the true multiplier is
actually with the agents and
>> even like you're saying about it's tough
for employees. I actually think that
there's a great opportunity for
employees right now which we'll touch on
later. But I think there's a there's an
amazing opportunity for employees that
take the initiative to learn this stuff.
You end up becoming like a 100x
employee. And it's like you can now like
one like go to market strategist can now
take on like the work of a 10 person
team and get paid for probably about
five five person team now.
>> Yeah. We're gonna dive deep into what I
think and what you think the future of a
team looks like. What are the necessary
skills? What does that brain need to
look like? And like I mean dude, we're
seeing like I think there there was a
company that was a billion dollars that
was one person. Like I truly believe
there's going to be one to three person
teams doing billions and billions.
>> Agreed. because there's so much of your
time and day that's like taken up with
just like stupid manual tasks. Like I
give you an example yesterday someone
from my team was asking for like some
passwords and API keys and it's like
before I would have had to go into one
password like find them create a share
link share it securely and like send the
Slack message. So instead I just got my
agent to just like go into one password
like create the share links send them on
Slack and it's like that would have been
maybe like a 10 minute task
realistically but it's like just 30
seconds and it's like it just allows you
as a founder to spend your time on the
work that's fulfilling for you and
creative and strategic the work that AI
like can't do like the visionary stuff.
>> Yeah, for sure. Let's dive into, you
know, what does that agent look like?
Let's really really deep dive into it.
For anybody out here that's listening, I
highly recommend that you either watch
this on YouTube or Spotify so you
actually can see the visuals, but let's
break down, you know, agent a bit more
and like what that kind of loop looks
like.
>> 100%. So the word agent gets thrown
around so much. It's kind of like lost a
lot of meaning now. You you see like if
you open Instagram or X, it's just I
build an agent that does this, agents
this, agents that. But the easiest way
to like actually think about like what
an agent is is I like to think chat is
just question to answer and an agent is
just goal to result. So question to
answer goal to result is just the simple
way to think about it in your head. So
for example with a chat you might ask it
a question it'll reply and then you have
to go and do the work. Whereas agents
you're giving it a goal it's breaking
that down to a plan and executing and
then handing you like a finished
product. And the word agent actually
just comes from the word agency which
means like ability to act and do things.
And people think like agent as in like
real estate agent or like travel agent
as like a person like a guy that's like
a sci-fi type guy that's like living his
own life and then doing work for you.
But in reality it's just like comes from
that word agency which was like aentic
AI which meant like AI that can do
things for you and then people just like
now call them agents. The way I think of
it is like chat models is like a really
brilliant consultant. Like you can go
ask it advice and it will like give you
give you the the answer but you have to
actually go and take that answer and do
the work. Whereas an agent's like hiring
that brilliant person in house that will
then go and do it for you too.
>> Yeah. End to end like you what do you
want? What does that result look like?
They figure out kind of the workback
schedule and then literally like when
you press press go and we're going to
talk about underneath the hood like what
it looks like to obviously the piping to
make that work. You just see it just
typing and working and it just gets the
end result.
>> Yeah. Because I would argue that chat
models question to answer like can't
actually even take any meaningful work
off your plate. Like maybe a research
task, but realistically like chat models
can't actually take work off your plate.
>> Yeah. Let's talk about like a couple of
those things. Let's like kind of
categorize like stuff that maybe does
make sense in chat. Like for me it'll be
just like, you know, I'm staying at this
hotel. I'm going to studio and then I'm
going to
book this hotel or this hotel and I'll
screenshot a bunch of them and it'll
just kind of optimize from a distance
perspective how I get to places or like
should I order this piece of food, this
piece of food like little simple things.
>> Yeah. Like yesterday I was on the subway
and I took a screenshot on Google Maps
and sent it to like Claude in just the
chat and I was like yo can I pay Apple
Pay on the subway cuz I'm from Australia
and it was just like yep you can. I'm
like sweet. Like just little questions
to answers,
>> little simple things, you still should
use chat.
>> Yeah. Yeah. Agreed. Cuz like sometimes
using an agent to ask like how many
calories in a banana is like killing a
fly with a bazooka. It's like overkill.
Like you don't need an agent for that.
So chat is still like very useful for
me. Like I just got called on my phone
and like asking questions like that
throughout the day. But any meaningful
work
>> straight to the agents.
>> Yeah. So, I mean, like a good way to
illustrate kind of this old way of
working with chats versus this new way
of working with agents is like a
marketing example because obviously like
I used to run an ad agency. It's kind of
used to be my bread and butter. And if I
was going to create a marketing
campaign, right, let's just say you're
like doing a Christmas in July campaign
for ecom brand. The old way was you'd
probably have the brief in notion or
whatever project management tool you're
using. You would copy that Christmas in
July brief, paste it into chat GBT. Yo,
we're doing this campaign. Like, create
me some ads. Write me the email copy.
You would then like take that copy and
like paste it into Clavio or then like
give it to your email designer. Get the
ads and then like go into some other
tool and make them. And you'd copy the
content from chat into your content
calendar, paste it, and then you would
have to go into like ads manager and
like manually input it all as well. Like
once
>> switching cost and a ton of like small
optimizations. It's it's going to two,
three, four, five different people.
>> And it's back to the thing about like
more fulfilling work as well. It's like
when I used to have to upload if you
were uploading like a 100 ads to ad
manager like the UI in meta ads manager
is so awful. It would take you like an
hour of like just data entry like
putting them into ad manager and
switching off all like the creative
enhancements and all the all the
[ __ ] So the new way of working
would literally be you just tell Claude
or whatever agent you're using
kick off the Christmas in July campaign
and it would go into notion. It would
find like the the campaign outline about
like the plan and like what specific
discount you're going to give and the
offer and stuff and it could go into
Shopify like pull in actual product
images. It have like your brand
guidelines so the email would be like
nice and branded and then you would like
write the ads. You can even use a tool
like Higsfield or an image gen to create
the static ads and then it might just
give you back like five emails and 100
ads and you pick like your favorite
email and like top 10 ads and then you
approve and then it'll just literally go
into Clavio, set it up and then go into
ads manager, set it up and like it's
done.
>> Yeah.
>> And and it could take it there and just
leave it as a draft and obviously you
can by hand optimize it before you click
go. And for everybody out there
listening and watching just to give you
context like we're talking very macro
right now. We are in this episode going
to actually show you live how to do
stuff like this because there's a lot of
there's a lot of stuff like in between
there to make sure that the piping is
set up properly to get that end result.
But yeah, man, I just feel like all of
like the little small work is going to
be absolutely done if and when you set
up the proper piping. It's really just
comes down to the proper piping and
systems that you have underneath all of
this.
>> Yeah. Yeah, pretty much. So, you've
given the agent a goal. It's it's done
this amazing work for you and given you
the output, but like what has it done?
like what happens in that mysterious
agent step where it's going and doing
the work. I think that it's really
really important to understand like
what's going on under the hood here. And
I'm just going to do a high level cuz
I'm I'm not a technical person myself
and I'm not going to lose anyone with
like going into like the actual software
engineering part of it. But I just want
to touch on this loop that agents go
through. So when you hand an agent a
task, it's literally just going through
three steps. observe, think, act over
and over and over again in a loop until
it has completed the task and then it
hands you the final result. So when I
say observe, think act, you can think of
it as look at the situation, decide the
next step, do it, look at the situation,
decide the next step, do it over and
over and over again.
>> Optimizing, optimizing, optimizing.
>> And then finally, it will output the
result when it determines that the task
is done. And then how it determines when
it should stop going in loops is based
on what you've told it in the goal. So
if you for example said like to an agent
as a simple example, research the top 10
business podcast I should go on and put
it in a 10page PowerPoint deck.
The agent will not stop going in loops
until it has 10 business podcasts in a
PowerPoint to give to you. So if you're
really vague with like what you want the
final output to be then um be a little
bit more subjective based on the agent.
It's not going to be exactly what you
want.
>> Yeah. Exactly. So, it's important.
That's one thing we'll touch on is
clearly defining like what done looks
>> But, you know, when you're giving it a
goal, it's just going in these the these
three steps. Look at situation, decide
the next step, do it.
>> And I thought what would be cool just to
like sort of illustrate this loop is to
just give a simple example into Claude
and I can show you this loop going in
real time. Before we we show that to
everybody out there, I think a key thing
to touch on just because guys, I want to
make sure that you're following along as
we go in that observe loop that we're
popping up right now. You see the
context and skills. We are going to get
very very deep into how to set up the
context and the value of the skills.
>> Yeah. Let me give this little example to
Claude so I can show you this loop in
action. Okay, so I'm in Claude Code
right now and we're going to just paste
this in.
>> Welcome back, Rem Dog.
Before you dive into this right now,
guys, this is going to be awesome. you
guys are going to see the whole entire
loop just so people have like context. I
don't even know if you even maybe want
to zoom out from a macro perspective or
we can do it after the fact of just like
driving around Claude and just telling
them like this is code, this is code,
this is chat. Let's just make sure that
we do that before the end of the
episode.
>> Yeah. Yeah, of course. And like I'm I
want this episode to be core principles
timeless. So I'm not going to go too
deep into like right now the best model
to use in cords opus 4.8 or Fable 5, but
in 2 months time that's going to be
different. But these core principles
that I'm going to cover today about like
how to make an agent actually useful and
work for you and like how to build
agents essentially that's not going to
change for months. So for this example
what I've told is I've said please give
me a report of my best performing
Instagram videos and a full breakdown of
the angle the hook topic title hooks and
why the video worked. I've given it my
Instagram handle and I've said create a
HTML report. So that's just going to
keep going until it's done an analysis
of my best performing videos. I could
have even give it a number like 10 best
performing and then it's going to
deliver the HTML report and that's when
it will stop the loop and give it back
to me. So
the first act step that it decided to do
was go and search my computer to see if
it can find more information about like
title hooks and hooks and some of the
things I've talked about. And it's then
found the title hooks are exactly what I
need. So now it's gone back to the
observe step. Now it's evaluating the
situation after the first loop. So
second loop round. Now it's going to go
and read. It's found some files. Now
it's evaluating the situation. Great.
I've found more files. What should I do
next? Let's look at these files. So you
can see it's read these two files. Back
to the observe step. Great. This is a
gold mine. There's a complete data set
of 60 reels matched with play counts.
And then it has I mean I think maybe I
have my video my real performance
already downloaded on my computer from
another project that it's found. If it
didn't find all these 60 reels with
data, it would have instead of that, it
would have gone on the loop and then
probably called the tool to go search
Instagram essentially. So, second loop
round, it's it's read the files. Now,
it's like, excellent, this is a gold
mine. It's read some more files. Another
loop. Now, it's gone. I have a complete
real data set. All 60 reels matched with
play counts, engagement rates, dates,
patent tags, all that stuff.
And then it said the one field that
genuinely isn't in this data is the
spoken hook first 3 to 5 seconds.
>> That's the bottleneck.
>> Yeah. So now it's gone. It's thinking
it's like well we need to go find that.
So now it's going to run a command and
it is going to go launch a sub agent
which we'll touch on later which isn't
as crazy sci-fi as it sounds. It's just
basically it's launching another version
of itself to just go check. And then
it's back in the observe step. Great.
Now we've got it. The data is rich and
complete. The spoken hook retrieval is
running. Let me build the full report.
And now it's running another act step to
build the report. So, you know, we've
got this loop here that the agent runs
through. And a application or or a
program that facilitates that agent loop
just allows that agent loop to run is
known as an agent harness. And then all
of like the top platforms you see online
now that you hear about like Claude
Code, Claude Co-work, Codeex, Manis,
Plexity Computer, those are all just
different agent harnesses. They're just
different applications to facilitate
this loop. So when people talk online
about I built an agent that does this,
that would imply they've built a
competitor to these like they've built
their own cord code. They're not
building
>> probably not.
>> Yeah. Yeah. They're not they're not
building a loop. Like the better word to
describe it is they tuned an agent. So
out of the box they all go through this
loop, right? But we can actually we have
four levers to pull to make these agents
better, which is kind of like the whole
concept of this episode. We're going to
just do a deep dive into each of these.
But at the observe step, that looks like
context, which is the information you
can give it, and then skills, which is
the processes that you've given it that
it can follow. On the think step is the
LLM model. So obviously these agents
that step where they're you know
deciding what to do the brain behind
that is the large language model which
you like the large language model is
like GPT 5.5 or opus 4.8 like the actual
model itself and then on the act step
which is giving it more tools. So like
out of the box like Claude Co-work or
Claude Code or most of these agents,
they have like a few tools. One of them
being like a web search, but if you want
to then have the agent like handle your
email inbox, you need to tune that act
step and give it more tools.
>> Yeah. I think what's really interesting
about that is it's almost like sports or
like F1. Like you could have a
Lamborghini and then you can have a
Volvo. Those Let's just look at those
and make the analogy that those are kind
of the LLM models.
>> Yeah. The Volvo could beat the
Lamborghini if the context, the skills,
and the tools are really [ __ ] tuned
and super super dialed.
>> If you put like a V12 engine in
>> I'm not a car guys. I don't know if that
was I don't know if that was the right
analogy, but it's just like
>> well like this this is like one of the
analogies I love to give is it's like
today in this episode I mean you've
already started but you're learning how
to drive.
>> Yeah. And all of these different tools
on the market like Manis, Perplexity,
Claude, Codeex, you can think of them as
all just different cars. Like Codex is a
Ferrari, like Claude, Claude Code is a
Range Rover, Manis is like a G Wagon.
And as long as you know how to drive,
which is like just the the core
fundamentals, which you know in driving,
it's like steering, brakes, accelerator,
you can pretty much hop in any of these
cars and within 5 minutes like figure
out how to work it. you know, like
there'll be slight differences. Like the
indicator might be on the other side or
like one might have another cool feature
like cruise control that another
doesn't, but essentially they're all
just the same thing and as long as you
know how to drive, you can pop open any
of these and and figure it out and start
using it efficiently. Like
>> hot take there. Everyone is just so
bullish on a specific LLM model, but
we're going to prove otherwise today.
Let's dive into what does that loop look
like right now on the IG side.
>> Let's check on this on this little
thing.
>> Let's check on old boy.
>> All right. All right. So, we can see we
got a nice report here, which is what
actually works on AI with Remy. And you
can see it's analyzed all 60 of my reels
with the top one being like over mill
views. And we can see that it's like
categorized them. It's like kind of put
them into into categories. It's done a
breakdown of each one with like the
title hook like on screen like
>> And if you guys look at this too, I
mean, you can take direct inspiration
from this. Like you see the right verse
wrong way to use Claude. You can do
another video. It says right verse wrong
way to do X. It's saying don't ask
Claude for business advice. Obviously
contrarian in nature. You could just say
don't ask X or don't do X to kind of you
could literally just take all of these
to make all of your new ideas.
>> Exactly. I mean this was just a report
to illustrate this example. Straight
away I can think of a use case here.
Like if you do content and brand deals
like you might want a report like that
generated for videos you did for the
brand. So they can have like a a report
of how the video performed for them as a
company. If you're doing brand deals and
stuff. There's so much stuff you can do
here. You can see it's just gone through
analyzed all my videos and like picked
them apart like why it worked, what
didn't. We can see here all 60 reels
ranked by plays with the date, title,
hooks, the durations, and then
throughout this process, right, like I
started to illustrate before how it's
going through loops about like, you
know, it's observing. It's decided to go
find some files. It's fed that back in.
Now it's decided like, all right, go do
this. It it it couldn't find some data
like the first spoken 5 seconds of the
videos. So then it had to like call like
make another couple loops to go find
that. And hard to say, but I reckon this
probably would have ran through about
like 20 or 30 loops before it decided
that the task was done. And you can see
here it worked for like 20 minutes or
whatever it was. So it just it will just
keep going until it's decided like, "All
right, this is done. I'm ready to give
it to Remy."
>> Yeah, it is amazing. And for anybody out
there listening, you know, if you're
running a company, if you want a report
at 7:30 a.m. in your mailbox every
single day, a beautiful PDF looking at
social media analytics, e-commerce
analytics, anything. Yeah, we're going
to walk through how to do this. But it's
crazy how granular this can get on, you
know, that individual prompt. Quick
second, Remy has shared so much in this
episode. All of his principles and
frameworks we put into a free PDF
playbook. You can download it below,
link it in the description. Enjoy the
rest of the episode. It's really about
what's what's underneath this and
obviously the piping. So let's get into
how to actually, you know, you call it
tune to to optimize the agent to make it
useful. So why don't you kind of walk us
through
>> those three big pillars and we're going
to deep dive into each.
>> 100%. So
>> a raw like agent out of the box like if
you open up Claude Code or or Co-work or
any of these popular agents on the
market and just like give it a task like
it's they're really capable. probably
not going to do it in the way you like
and you're probably going to get an
output you're not happy with. And that's
why a lot of people like start using
these tools because it's all the rage
and then they just quit because it's
like, oh, it didn't work for me. In
order to actually make the most out of
these agents and make them actually
useful, you need to treat setting them
up like onboarding a real employee. So,
you can think of agents as like really
really capable strangers. Picture that
dude, you know, like we've all got that
friend that's like really really smart,
like good at solving problems, and we'll
just like keep working until like the
task is done. you now have the
opportunity to hire this guy into your
business for $100 a month in salary,
>> $1,200 a year. Sign me up.
>> And and he has infinite bandwidth. He
can just like like there's no limit to
what he can do. But no matter how great
he is and smart and capable, unless he
fully understands you in your business
or if you're an employee, your role and
what you do, unless he has the tools he
needs to actually do the work, like an
email account, a browser, a computer,
and lastly, unless he's trained on like
your secret source that makes your
business unique and your processes, he's
not going to be very useful at all.
>> All right, so let's actually onboard
this hire. And you mentioned context,
tools, and skills. give us full context
of how you look at onboarding the hire.
>> Like I said before, this this really
capable stranger you've now got working
for you, he's not going to be able to do
very much unless he fully understands
you and your business, has the tools to
do the work, and then knows how you like
the work done. And you do this through
context, tools, and skills. So context
being like information to the agent,
like what you do, your business, like
your brand voice, things like that.
tools being like, you know, a computer,
a browser, Gmail, notion, Slack, and
then skills being like your SOPs, like
your special source, like literally like
documents like you give a VA. And then
if you get those three things right,
that's how you like properly on board
the stranger, and then you have the best
employee you ever had.
>> Sounds amazing. Let's go super super
deep into each one of them. So, we have
context, tools, and skills. Let's double
click on context and tell us exactly
what you mean by context. Yeah. So,
context is literally like everything
that the agent should already know
before you give it a task. And for me,
this looks like things like about me.
So, all information about like me as the
user cuz I want them to like I want my
my new employee to understand like
everything about like who I am and what
I do. Even some things about like my
story and like my come up and things cuz
then it can use it if it's ever writing
like a landing page copy or like a sales
letter script. it can reference like
real things in my life. It also looks
like a document about your business
information. A like a maybe another
document of like your offer catalog
where you talk about like all your
services and offers and their prices and
how they link together like upsells,
downells. Another one might be like an
ideal customer profile which is one of
my like most important ones. So it like
knows who your business like targets and
quite literally
like if you were onboarding this new
employee, right? like a real employee,
you'd probably like give them a PDFs or
like some sort of documents, usually
like Google Docs, Word Docs that teach
them all of this stuff or like the key
information you want them to know. You
you are literally doing the same to the
agent. So like just like you know you
get get a new VA, you like give them all
these docs, you're literally giving
these docs to the agent and these are in
the form of markdown. So for the
listeners that don't really understand
what markdown is, markdown's like it's
it's basically one step above a .txt
file. So TXT text file is just literally
plain text. No headings, no bolds, like
nothing. It's just text. Markdown is
literally just one step above that. It's
just a plain text file, but you can do
like headings, boldings, like bullet
points,
>> simple formatting, stuff like that.
>> Yeah, it's just like a slightly more
formatted text file. And you know, if if
any of you use out there use like a tool
like Notion, the notion docs or like
ClickUp, the ClickUp docs, they're all
in markdown format. Just super simple
like headings one, twos, and threes,
text, bullets, and bolding and like
italics. Just the super basic
formatting.
>> And to give you guys further context,
it's basically like a simpler doc file.
So, it's a text doc. It's somewhere in
that universe.
>> Exactly. And the reason like why
markdown is so important is you could
theoretically let's just say you're
building you're following along and
you're trying to build out your own
agent. You could just give them all
these in a docs file, right? Or a PDF.
However, it it's a lot harder for the
agent to access the information. Like if
you give an agent your ideal customer
profile as a markdown file versus a PDF
with the PDF, it's going to have to
unpack it and like try and extract the
text cuz it's just the way that the PDF
file is like built. It's hard for the
agent to like ingest the information.
Whereas with markdown, it's just like
that straight ingest. It can just read
it really really easily. And we want to
make things easy for the agent because
that just means it's things are quicker
for you and things are cheaper in terms
of like cost for the model running. So
markdown is just the best way to be
giving information to your agent.
>> To give people context as well, how many
MD files do you have on your computer?
>> Like hundreds.
>> Hundreds.
>> But like so the these you're literally
building out like markdown files for key
areas of like information that you want
to give this agent.
>> And as far as that goes too, you can get
would you just say the more granular the
better?
>> Not necessarily. I don't really think
there's like a right or or wrong way to
do it. It's more like I I just build the
markdown files for me almost like how I
would want them formatted. Like I've
just like literally got like a about me,
like a business info, brand voice, idle,
customer profile, like offer catalog. I
could put them all in one doc, but then
it's just like I don't know. I almost
just prefer them having having them
separate.
>> Yeah, we're going to get into folder
structure as well, too, so everybody
understands kind of from a macro
perspective how this brain looks and how
it kind of all flows. I think for the
people out there listening, you know,
we'll get in, you know, what is the
first thing that you can do to kind of
start your journey. I feel like like the
about MD file is like a pretty clear one
of like
>> great spot to start. And the thing you
need to understand is these markdown
files you're building are like assets.
They're assets for you and your company.
So like really like spending the time to
build these docs like really really
well. Like if you just like take the
time to build a really wellthoughtout
about your business doc and that like
the ideal customer profile they're
assets for your company because if you
rush the ideal customer profile doc and
it's just like a bunch of AI slop that
you got chat here to put together and
you haven't read over every line. There
might be a line in there that says like
my ideal customer like if you're a
skincare brand a line in there in your
ideal customer profile might say their
problem point is like red skin or
something. And if you don't check over
that now in all your marketing copy,
you're gonna have like it targeting
towards the angle of red skin. And if
that's not something that your product
solves, then that's like a that's an
issue. You know,
>> the answer on the MD files is probably
you want to minimize the subjectivity in
the MD files. Make them all objective,
like stuff that's kind of black and
white. Like I definitely would use this
as an input to make decisions.
>> Yeah. Yeah. Agreed. And I and I get
Claude to build all these files for me,
but I read over every line before I like
give it the okay.
>> Guys, I got a little cheat code for you
too on the about.md file. Like the
craziest thing about all this is like
I'm not saying have Claude do it, but
you can say to Claude, I want to build
an about.md file. How should I, you
know, how should I process this? How
should I think through this? Yeah.
>> And that's where you I do think you can
use something like chat because you can
kind of go back and forth and build it.
>> Yeah, that's like a a perfectly fine use
case for like the chat models. Like
that's how I built out all of these. I
literally said to Claude, like in a new
chat, I was like, "Yo, I wanted you to
build out a bunch of markdown files for
me, like an about me, and I had a custom
profile, just listed out all the ones I
wanted, and I just said, interview me,
like to extract the information from my
brain to build these files, and I sat
there like using Whisper Flow answering
all the questions for like 3 hours,
>> just riffing, and then Whisper Flow just
cleans it all up."
>> Yeah. It just like sucks it all out of
your brain into these files that are now
assets for your company that you can
then give to your agents. And like one
of the key distingu like the key
distinguishments to make is when you're
using these chat models, they're sort of
like a black box. Like everyone always
goes on about like my chat GBT knows so
much about me. I don't want to move into
Claude because like Claude doesn't know
me. ChatGBT knows me so well. And it's
like, you're right, but what it's doing
is it's like got pretty much just a
markdown file like memory MD or like
about user.md in the background and
chatgbt is building this file on your
behalf without you knowing and you can't
access this file. So if you're like
starting a chat GBT chat and you say in
there like I'm I've just moved to New
York City like what are some things like
good cafes near me? JGBC in the
background will take out that snippet of
like I just moved to New York and save
it in its markdown file and then now
like it knows you've moved to New York
and you just think that this model's
like almost trained on you now but in
reality it's just a markdown file where
it's choosing what to save and the other
thing about that is you can't control
what it saves as well like your chat GBT
like knows everything about you and then
like you know you're writing like a
landing page you're asking for advice on
chat GBC like hey how do I break up with
my girlfriend and then you're writing a
landing page and it's like all of a
sudden the landing page copies about you
breaking up with your girlfriend. It's
like you can't you have no control with
these chat models about like what it
knows and what it saves
>> like it it does get mixed up sometimes.
>> Exactly. So you want to make sure that
you have full control. You want to own
that context and that's like I'm talking
about just in the format of these
markdown files and then you can they're
assets that you own now and you just
bring them to whatever agent you're
using. So like I've got my folder set up
which we're going to get into with all
the context in there and I can just open
up that folder in claude code or codeex
or manus and like straight away it knows
everything about me. It's like it
doesn't matter like I can just switch
and stay sort of like mobile. Guys, real
quick, everything that Remy is teaching
you in this video is absolutely free. No
gatekeeping as always. But for the
people who want it implemented, Remy
made a plugin that builds this entire
system for you around your business and
your life. The loop, the folder
structure, all of it. I am not
technical. I did it myself. 90 minutes,
one sitting, done. Sign up for the
newsletter below and get 50% off at
launch. Enjoy the rest of the episode.
So, the first step for onboarding agent
is like building out these context
files, right? And I'm going to show you
like what this actually looks like. So,
I'm going to open up Cloud Code again.
Here,
we're going to do a new session.
Nice and clean.
>> Hey, there we go. And I'm got my OS
folder open now, which we'll touch on
later. But the way that these agents
work now is a lot of them run like
locally on your computer as opposed to
like the cloud. And what locally on your
computer means is it's like instead of
having like a project on your chatbt
account, your that project folder now
lives on your computer. And there's a
lot of benefits to that which I'll I'll
touch on later, but I'm just going to
open up this one I've created now called
demo. So this is just a brand new
folder. You can see it's completely
empty
>> and that's just you just created an
empty folder on your desktop.
>> Yep. So it's desktop empty folder called
demo. And we're just going to open that
up. So now you can see in cloud code
we've scoped the agent to just work in
that folder. Just like you know when you
have a a project in chat GBT, it's like
any of the chats you do in that project.
It's like scoping chat GBT to just work
there.
>> So now it's a new blank canvas and
you're you're deciding what inputs go
into this folder.
>> Exactly. Exactly. So you can see just to
demonstrate this the importance of
context. If I'm opening up this brand
new folder, so I've been working with
claude code, right, for for like eight
months and if I ask like you know who am
I?
So you can see like when I've asked it,
there's no markdown files for it to look
at. So, I've asked it who am I? I've
been using CL code for eight months. It
should know. But it doesn't. It's
because all the context you own in those
files. It's not a black box.
>> So, if someone has an alter alias, they
can they can just create different
alteres in in in different folders.
>> Yeah. Essentially, so you can see here
it it tried to read the memory MD mark
file, markdown file, but it doesn't
exist. So, you can see here it's just
like the f file doesn't exist.
>> So, it went through that loop. It was
searching, but there was no kind of raw
data to extract.
>> There was nothing for it to to do.
>> So, how do they know the email and the
likely name? because that was just from
my claude account. It must just have
access to that. That's literally just
because that's the email associated with
my claude account and name. Actually,
no, no, no. Actually, it's it's pulled
that from my computer because this agent
lives on, you know, it's got you got a
username on the computer. That's where
it's gotten it from.
>> Okay. So, what does this look like when
you're on steroids? Let's dive deeper
into kind of doing these MD files.
>> Yeah. So, let's let's continue
onboarding our agent, right? So, the
most important markdown file, like, you
know, we've got our markdown context
files. The most important one when
you're dealing with agents is what's
called an agents.mmd file. And I mean
claude have just decided to be difficult
and in in their systems they call it a
claw.md. But in pretty much all the
other harnesses you use like codeex and
open claw and stuff it's called an
agents.mmd. And essentially what happens
is is every time you're like starting a
new session and by new session I just
mean like if we go back to that goal to
result thing it's like every time you're
going to like give it a new goal you're
kind of going to click like new session.
the same way you start a new chat in
chat GBT.
>> Each session has an amount of like
space. It's kind of like you can think
of like the agents like brain power and
as the session starts to like fill up
the you have what's called context rot
which is basically like have you ever
been in a in a chat on chat GBT and you
it's like so long you've been in there
for hours and it starts to just like go
crazy like it literally just starts to
just like hallucinate forget things. So
what what's happening there is every
time you start a new session like a
fresh like chat window I guess it's got
like a certain amount of of I guess
capacity you could call it like free for
the actual work and what your sessions
probably look like is something like
this where they're just like like so
like so much information that the stuff
you started telling it at the start like
fall out the top
>> that was V1 Mark bro I'm not on V1 Mark
anyone I mean I think 99.999% of the
world is in probably V1 Mark one endless
chat to be frank with you.
>> Yeah. Yeah, probably. So, every time you
start a fresh session, the way that this
agent like knows about you. Same way
that chat gbt's memory works is if you
start a new chat in the background, it
started the chat with that memory. MD
loaded into it, but you can't see that,
right? So, you give it a prompt like uh
write me an email in chatgbt and to you
it just shows as like that one prompt
you've given chatg, but in the
background it's also loaded in that
memory markdown file. So instead you've
got like like a hundred lines of
information about Mark and then write me
an email. So it's got all that context
in it. Same thing happens with agents,
right? So we've started a fresh session
here. What this actually looks like is
any file, any markdown file in that
folder titled claude.md gets
automatically loaded in. So your
claude.md or agents.mmd
that's like your your northstar markdown
file. That's the one you need in every
single project you're working in. So to
illustrate my point, I've created this
folder here with a claude.mmd in it,
right? That I've pre-prepared. And we
can actually just open this up. So you
can actually see what a markdown file
looks like. And I got Claude to build
this for me. So we can see Rey's AI
operating system. Who am I? Rey's the
founder of AI with Remy. The business,
the brand voice, the ideal customer, the
offer, the values, how I like to work,
preferences.
This is like our you know northstar
context file and that's the one that's
going to load in to each new session by
default
>> to give macro context on how you're
thinking.
>> Yeah. To give macro context essentially.
So that same example right if we go back
into claude here we just asked it in
that folder when it was empty who am I?
It had no idea. Now if we go new session
same folder but it's now got that
claude.mmd in it. And we can say who am
>> Heyo.
And you can see here now it knows
exactly who I am. It knows exactly like
what I do, my background, the brand, the
three products.
So that's like a great illustration
there. Do you see how how it goes from
knowing nothing about you to now having
that context file that knows everything?
>> That's why it's so important to be lean
and intentional on what files you upload
obviously in there.
>> Yeah. Yeah. Exactly. So that's that's
your pinnacle markdown context file.
That's like the first one that you need
to build. the the part about the
claw.mmd made sense, right? That gets
slotted in and it's like the agent wakes
up to do a task and he's got that loaded
into his brain, right?
>> Yeah. And that's something that's you
should constantly be updating if
something big in your life changes.
Correct.
>> Yeah. Yeah. Correct.
>> Yep.
>> So, you're probably like wondering now,
okay, great. I get that claw.md
agents.mmd. Where do the other ones fit
in? Like the about me and the customer
profile, like where do I put those? So
those ones, the other markdown files,
they don't get loaded into that session
window by default. So we need to we've
got our CLMD that gets loaded in by
default. And then we need to tell it in
that that it will always read, hey, go
check here as well. So then what I like
to do, how I have this set up is we got
our claw. MD here, and I'm going to
paste in these other ones I've
pre-prepared.
So now you can see we've got a couple of
other files taking shape here. So we've
got our claw.mmd.
We've got this context folder with all
these markdown files in it. My values,
my offer, my item customer profile,
brand voice, about me, and then this
memory.mmd, which we'll touch on in a
second. So by default, claude is any any
project you open up claude in to work,
it's always going to read that
claude.mmd. And then we're going to
point to the context folder in the
claw.md. So you can start to like
compartmentalize and like modulate your
your project, if you will. making like a
whole like a like decision rubric like a
decision tree of you know if it goes
here it opens up this folder and stuff
like that.
>> Exactly. And then this is like all these
agents you build they're super like
modular. It's like you hear other ones
really popular like Obsidian is a really
popular thing that people give to their
agents as like a second brain
essentially like there's loads of
different like frameworks like that. And
instead what they're doing like if if
you had this folder and you wanted to
use obsidian instead you would instead
delete this context folder and you would
in your cloudmd point to the obsidian
vault instead. That's kind of how people
are doing that like you can kind of
modulate the the context.
>> So in short you can kind of redirect it
whichever route you want to go.
>> Yeah. Exactly. And it all stems from
this file here. So now this new setup
we've got is kind of like evolved. So we
just the first one was just the
claw.mmd. And that's probably the most
basic one that if you're trying to like
start using agents for yourself. That's
probably the first step I'd say is just
create a folder and then just build a
cloud. Mmd just ask claude, hey build
this claw.mmd file for me and just
include in it like an overview on who I
am and what I do just to give it like
that base context. Then v2 is now we're
adding in like this extra folder for
like we're expanding out our our context
here. And if we open claw.md now, now
I'm going to point to that context
folder. So you can see if we open it up,
it's like who am I? My context read
before any task. Everything you should
know about me and the business lives in
the context folder and then it's just
given an overview on the files.
>> Wow.
>> Load the relevant files in before
starting any tasks. Assumptions are the
enemy. If the answer isn't there, ask
me. And then everything on continues
like normal. And then the memory I'll
touch on in a sec.
>> And then as you keep developing more
folders, you just keep updating. The
most important thing is the claw empty
file cuz that is going to create the
infrastructure like you could have like
marketing, finance,
>> logistics.
>> And then underneath that you can have
multiple folders.
>> Exactly. So when I asked Claude before
like who am I? It knew straight away.
That's cuz that file automatically
loaded in. But if I deleted that claw.md
and then asked it who am I again?
>> Back to dummy. Back to dummy.
>> Well, it's got the context folder.
>> Oh yeah.
>> If I just deleted the CL MD and left the
context one, it wouldn't know straight
away. It would then go through a loop to
think like, oh well, I don't I don't
know who who he is. Uh, let me think. I
can let me let me check for what folders
I've got. Oh, context. Then it would
read it. Then it would load it in. But
it's just like it's slow and it's got to
go through another loop. So like that's
why you have the claws that MD there.
>> Why why don't you everybody listening
right now just ask it something very
very basic under the notion that now you
have context based in like something
about any of those kind of raw folders.
I know you had under there like brand
voice like anything.
>> Yeah. Yeah. Well, what I what I might do
is I'll just say like write me a cold
email to
>> how about to Mark to come on uh to come
on open residency.
>> Actually, know this. I'll just say like
>> I'm going to whisper flow it to demo.
>> Yeah. Amazing.
>> Makes sense.
>> Write me a cold email to Mark from Open
Residency podcast to try and get me a
position on the show.
>> Wow. Whisper for the win, baby.
>> Whisper flow is so good. Like talk about
like agents stacking up the
productivity. You can even stack on top
of the agents when you start using
Whisper Flow because it's like you can
just just get things down onto the page
at the speed that you can think.
>> For the people out there listening, I'm
going to give you guys a little cheat
code. One of our episodes a bit before
this is of Chris Voss, which is a
negotiation expert.
>> You can make some MD files with him to
speak in his language. So it probably
would be a bit more persuasive on this
email that you're going to send me.
Let's let's let's see how good it is.
Let's see your sales pitch right now.
>> Okay, great. So it's gone through a
couple loops, right? It's got decided to
go and research who you are. It's gone
to go and like find
>> that is me
>> open residency
>> and now it's written right. It's like
literally said here like like we haven't
given any other context. We just said
write the cold email and it's got cuz
it's got our our clawd and then this
context folder and that context folder
is so rich with like my values, my
philosophies, my like thought processes
on things and my offers. So you can see
like the subject lines, right? Operating
a 20 person team without hiring 20
people. guest idea for open residency.
But the email right here, hey Mark, been
working through open residency. The
operator as last line of defense framing
is the realist take on running a
business. Terrible cold email by the
way. Terrible. But it's basically you
can see here it knows everything about
me. It's like I teach founders to stop
drowning in tools.
>> This is where we're going to get into
like the additive layers of like an
example is like that negotiation
>> skill that you know how you speak skill.
Like you can get way deeper to make this
thing absolutely flawless. Clearly, it
did an amazing job.
>> Yeah. I mean, the key point demonstrate
was right that it just knows about me.
It it can write an email with all the
context. Like, it doesn't like the
email's terrible, but like and you can
tell it knows about like what I do and
my philosophy.
>> Amazing. You're booked. Let's go.
[laughter] We on. We're here.
>> Okay. Great. So,
>> and just to take a step further, just to
kind of paint the picture of where we're
going to go with this.
>> This could be this email is drafted. It
could also be automatically sent as well
too, which we're going to get to to
connectors and MCPS as well too, just so
people know like this could be fully a
toz. Find all the business podcasts,
draft the email, and send the email.
>> Then you schedule it. So every week it's
going to or every morning it's going to
go find business podcast, cold email
them to get you book calls, like all
this stuff. It's crazy.
>> You're going to need to dial that email
in a little better for me to say yes
though, but it's okay. Don't worry. Keep
going.
>> So So that's that's what I'm going to do
now. I want to display like so
understand like we need these context
markdown files. The most important one's
the claw.md that then points to the
others. And now like when we open up a
new session, it it it straight away
knows everything about us, right? But
the next layer of context is memory,
which is still kind of under the
umbrella of context, right? Memory and
lessons. So off the rip, it knows all
this stuff about us. But what if we ask
it to do something differently? So you
can see I've got in here this memory.mmd
file, right? If we open this up, this is
just this has got memory learned rules,
lessons Remy has taught me. I check
these before every task and it's got
never write formally, just a couple of
like random preferences. So the agent
like knows everything about us out of
the box, but we want it to be like be
able to improve over time and like if we
correct it, we wanted to save that
somewhere, which also counts as context,
right? So I like to build this little
like kind of de facto memory system. I
think like when I started putting this
together, it was because none of these
agent harnesses had memory built in. So
if you said to it like right this this
email right here, if we said like can
you please sign off the email with
cheers Remy and like I always want you
to sign off emails with cheers.
>> It's not auto updating unless you put
>> Exactly. It's not so like before the
last maybe in the last month they've a
lot of these harnesses have now added in
memory as a feature but before that you
had to do it yourself cuz otherwise it
would say like if you didn't have that
in place it would say like yep got it
but then next session in a week time
when you write an email it wouldn't do
it.
>> Show them a memory MD just in case
they're driving the Kia and not the
Lamborghini and maybe that LLM does not
have it built in. Show them a little bit
of context.
>> So we can just show in here like the
memory
>> MD. No shame on the Kia either. I've
seen some nice Kia by the way too.
[laughter]
So, we got the the memory learn rules.
These just a couple I put in here
before, like never write formally, like
always use real examples, never generic
placeholder ones. And then same way that
we're pointing Claude at the context
folder, we're also going to point at the
memory folder. And I've just basically
said, you know, your memory lives in
this file. When I correct you, write
lessons into this memory file as a new
rule and then check it before every
task. This is how you go from good
assistant to know exactly how I like
things done. And I have one little tip
is if you use the at@ symbol if you tag
a file like that then it like it's
almost like a a sim link. I don't know
if you know what that is on a desktop.
It's like
>> I don't
>> if we for example if I tell it
>> is it like a lighter load on the brain
like that it zips right into the file or
something?
>> Yep. Zips right in the file. It's it's
like as good as the file being loaded in
if that makes sense. Cuz if we just tell
it like go use the memory MD and don't
add it'll then have to make a loop to go
and find the file and bring it in. But
if you just have like the app, it's
automatically loaded straight in.
>> Guys, we're going Lamborghini full
steroids here.
>> Hell yeah. So
Claude like has this this built in, but
even still, it's still almost a little
bit of a black box. It will save it to
like Claude's desktop app. So now if
you're trying to work with this folder
in like codecs where you want to try and
like use other harnesses, you don't own
the context file. Claude kind of owns
it. So like that's still why in all my
projects I still kind of add this memory
system in. So we can just test this out
here, right? And we can say
please in future when creating websites
I want you to always use dark mode and
never ever use light mode.
>> Here we go. It's done it. Bang.
>> Okay, great. So you can see it's now
gone and said add I'll add this to your
memory rules so I never forget it. It's
read the memory md and then you can see
it's added it's edited and it's added
this line. When building any websites,
always use dark mode. Never use light
mode.
>> Love that. That's obviously in that auto
updated in that folder that you showed
us.
>> Yep. So, if we go back here into this
demo folder, we go into our memory.mmd
file and open that bad boy up. You can
see when building any website, always
use dark mode, never use light mode. And
then now this is like we own that that
preferences, right? This will grow over
time. And like I've had the question
before, it's like what happens if that
just keeps growing? you've got like
hundreds of different lines in there and
they all start to like contradict and
get messy. Like I don't think that will
happen. Maybe in like 6 months time you
can just do a spring clean. Just
>> open a session with Claude and go like
>> just audit my memories and just take out
anything that's contradictory or
conflicting and cleans things up.
>> So that is pretty much like context 101.
Just to recap, we've got the claw.md
which Claude autoloads in every session.
We want to give it more context. we add
that extra little file and then point
claude to that folder and then we've set
up our little memory like
self-improvement system as well.
>> So from a loop perspective we're going
to start with the about.mme the claw.md
is going to be kind of the the macro
infrastructure which is then going to
lead into more folders which we'll
obviously get into and then the memory
MD is just a self-fulfilling loop that
just keeps remembering all the
preferences and such.
>> Yeah. Yeah. Exactly. So now like every
every session
we start that observe step now has all
the context and skills loaded in. So
every time we're giving it a goal
straight away in the first observe step
before it does anything we've already
got like now I mean skills will touch on
but now you can see we've really tuned
up that context step. So when it starts
to work on your tasks it has like all
the information that it needs. And I
think one last thing just to just to
finally like drive home this point. We
can run in Claude code this little
command. Don't worry about all the
others. We'll touch on that later. But
this is a one that comes built in with
Claude. You can go slash context and it
will actually pull up that real bar
chart for you. So you can see here
context window. We've used 6% of that
bar.
>> Are you talking about 6% of like the
memory within that individual chat?
>> Yeah, within that individual chat. So
you can see it's already like our bar.
used up 6%. And if we look at like we
we've given it one task here, right?
Like please in future tiny little
message and it's 6% which is a fair bit.
It's almost like it's a fair bit of of
the total space used. So you're starting
each session even if we just did like a
blank thing with no messages sent you'll
always like have a little bit of your
window chewed up because it's feeding in
all that context before you start any
task. That's why you could also argue
going back to the beginning is keep the
MD files like lean and cut, you know,
only use what you need to use.
>> Yeah, exactly. Because if you have these
like really bloated context files and
like you have, let's say you have 100
markdown files, you start a session and
the context window is already like 20%
full. It's like you you want to Yeah.
100% to your point, you want to keep
them nice and lean and like really like
potent with like high density
information.
>> Amazing.
>> I use Whisper Flow for hours every
single day. I think way faster than
[music] I type. I speak about 220 words
a minute and type only 45. Whisper Flow
closes that gap. It's a voicetoext tool
that works anywhere on your computer or
phone. You talk and clean formatted text
shows up wherever your cursor is. All
right, let me show you. I'm going to
answer a real email right now without
typing. You can see here Tiger is asking
for feedback on our intro video for the
next episode. Watch how I respond. if
you could just make a few changes. Uh,
make the red a little bit brighter,
[music] the typography a little bit
smaller. Uh, for the music, if you can
just give me some other options, and
[music] that first photo, some more
options as well. Uh, please give me
three options for each. And, uh, make
sure you send it to me by end of the day
Friday. Thanks.
So, as you can tell, the text is way
cleaner. [music]
I kind of speak in half sentences and
sometimes mumble on. This makes it very
clean, formatted perfectly, and avoids
all the small mistakes. This is a
product that I cannot live without. I
promise you that you will love it. If
you guys want to get one month free of
Whisper Flow Pro, click the link below
and use code open residency.
So, yeah, context. We've we've we've
taken our really capable stranger that
we've employed and we've told him
everything he needs to know about us.
He's ready to start working for us, but
he's still going to be pretty useless if
you don't give him like tools to work
with. Like imagine like you've just got
this this dude that's working for you,
but it's like he doesn't have an email
account or a computer or a browser or a
phone. He's like still going to be
useless. So we need to give him tools to
be able to actually complete the work
for us. So the way we do this is what's
called a thing called MCP. It stands for
model context protocol. And I don't want
to lose anyone here. It's it sounds
technical, but like it's it's super
simple. And and this all MCP is is just
a way for your agent to connect to
tools, things like
>> third party tools.
>> Yeah. It's a third party tools. Yeah,
exactly. Things like, you know, Slack,
Gmail, Calendar, like you need a way for
your agent to to connect and you do that
through what's called MCP.
>> To give people more context as well, I
think a lot of people out there use
either N8 or Zapier is a big one that
people use to kind of extract
information and and connect it.
>> And we're in the era now. I'm so curious
to see. I think all the principles here
are evergreen,
>> but it's like it's a big deal now to
have an MCP. And like some of these big
companies don't have MCPs
>> and like it's it's they're crazy. We're
going to get into the BI of MCP
>> that it's what I make purchasing
decisions. Now, if if I'm choosing,
let's say I'm choosing like a marketing
email marketing software, for example,
if like if I've got five options and if
if they don't have an MCP for a way for
me to connect my agent to it, I it's out
of the question. I won't use it. I don't
use front ends anymore of apps. Yeah.
like like the future of this is and
everyone that I know that works the way
this new way of working this stage two
which is agents they don't really use
like frontends anymore like I complete a
full workday from with inside like my
cloud code or or agent harness that I'm
using
>> by front ends just so people out there
listening like you're not in platform in
behind on the dashboard you're just
extracting from there and then working
within your workspace
>> y like I use Google calendar I use Gmail
I use Slack I use Stripe for payments
cal.com for bookings And I can't even
tell you the last time I've opened any
of those apps up. They're just all
connected into CLA. It's this new it's
it's a completely new way of working
where you've got
>> Claude or whatever AI harness you're
using that sits on this as like an
intelligence layer on top of all your
tools and then you're just interacting
with like this this intelligence layer.
So, you know, I've alluded to MCPS being
just this this way for your agent to
connect to tools. But super quick on
like what it actually is, right? So
before MCPs existed, if you wanted to
connect your agent like Claude to a
tool, it was it was difficult. It was
possible, but required lots of custom
development. And the way to think about
it is like Claude speaks English, right?
And all these different tools speak
different languages. Like Gmail speaks
French, Notion speaks Spanish, Slack
speaks Chinese. So if you want Claude to
talk to to Gmail to like get it to do
things, it was required a lot of custom
development. It required time and effort
because Claude had to learn
>> heavy, slow, expensive, everything.
>> French.
>> Yep. And shout out Ross Mike. This is
explanations from him. He was on like
some part I heard him. So this is taken
directly from him. This is what landed
like really nicely in my brain. MCPS are
just like this translator that sits in
the middle between Claude and your and
your tools. So Claude does not need to
learn French and Spanish and Chinese. It
just has this little translator in the
middle. So it just makes calls and then
it can be relayed really really nicely
to the tools. This is an
oversimplification but quite frankly
like no one really needs to know what
happens here. All you need to understand
is the MCPs are just your way to connect
tools into your agent.
>> Uniformity, simplicity, and that just
streamlines everything for speed as
well.
>> Same way you can think of it like an
API, right? For those of you who know
what an API is, like it's just a way for
tools to talk to each other.
>> And again, that's an oversimplification
of explaining MCPS and connectors. And I
think the big thing to take out of this
episode is the ability that you can ask
the LLM like, oh, how do I set up
connectors and how I do it? And I don't
want to encourage people to be lazy, but
you want to just quickly just show over
the screen of just where you go into
Claude to do it because it's very very
simple.
>> It's so simple. So the way like if we
open the Claude desktop app
>> and we go into customize, you can see
straight away connectors.
>> Boom.
>> And we have like we go up here to
browse.
There is so many connectors here. Like
pretty much all of the major tools that
you would use have a connection here.
And it's as simple as this, right? So
let's just say we want to connect up
ClickUp. So I can literally just go
ahead and go plus and then I'll just
oneclick login to my ClickUp.
>> Guys, for context, we are power users of
ClickUp. That is the harness the project
management tool that we use that creates
all the workflows, all the dependencies.
>> So we can go in here. Literally, I've
selected my workspace connect
open claude. It's connected. We go back
to the desktop connected to ClickUp. And
you can see here like we've got all of
these different things that we can do in
ClickUp. Create tasks, create lists.
Claude can now pull information from
ClickUp if you're planning like
marketing campaigns. It can update it.
And we've got all these like by default
set to needs approval. So it's always
going to ask you before it takes
actions. But as you start to build up
trust with Claude and like
>> you know you you get more comfortable
with allowing it to have more access.
You can just do always allow. Like all
my tools I just have always allow pretty
much. But in Claude like that's as
simple as it is to connect these. But
that is using MCP by the way. So you
don't really have to deal with any of
the technical [ __ ] But that when you
click the plus and log in that is it's
it's got an MCP server basically. That's
what it's using.
>> It's crazy that you say that like you're
picking vendors based on if they have an
MCP or not. It's literally like 5 to
10x. We're going to dive deeper on a
specific use case onto how much more
granular and strategic and faster and
optimized it is with this.
>> Yeah. I mean, it begs the question of
just like,
are a couple LLMs going to win and then
just the world is just going to be a ton
of MCPS and connectors into it?
>> Like, if you're building a software
product, it's like I I had a buddy come
to me and he's like, "Yo, I'm building
this like travel agent like AI agent
that helps people like plan their trips
and he's got this like nice website
where people log on and you you say,
"Okay, I'm going to New York for 10
days. Like, help me do this." And like
has an agent that plans your trip. Like
I I said to him, I'm like, I I don't
think that's the right way you should be
building products for the future. I'm
like, you'd be better off building this
travel agent as an MCP that you then get
hooked up into Claude and then now
people who use Claude can now have you
show up as a connector here, click plus,
your, you know, travel AI and then add
your product as an MCP and then now you
build a purposebuilt MCP for Claude that
helps that makes Claude be able to plan
your trips better that helps Claude like
gives it some extra tools it needs like
it can search, you know, like Sky
Scanner and some other websites a bit
easier.
>> Yeah. I mean, I feel like Yeah. I feel
like the future is like as opposed to
offering like a hundred things, it's
like be really really dialed on 10 20
things, then it becomes almost like an
external data play in conjunction with
that vertical.
>> Like,
>> yeah,
>> be really really smart and understand
and know less things and just feed it
>> into Claude. I completely agree. We talk
all the time about like mid-market B2B
SAS is absolutely smashed. like some of
these companies that got sold like I
don't even want to call some of them out
but like basically what I was doing on
>> these SAS tools
>> I'm literally just doing it inside
Claude right now.
>> Yeah. Yeah. Exactly. I mean like
>> Yeah. The the the SAS market is just
getting absolutely decimated. And I
think another thing is like a lot of
these software is built on like
headcount. Like they're built on like
seats in the software and now it's like
you just people are like reducing
headcount. They got agents in there and
like they can't bill for like agents.
That's bad. Yeah. So, you mentioned kind
of the always allow or kind of ask
first. Like that begs to ask the
question of like security. A lot of
people are very concerned about like
connecting a lot of this stuff. Like
what's your just overall thesis on how
people should approach security?
>> You know,
you just have to weigh up the trade-off.
And for me, it's like, yeah, there's a
risk of things going wrong. Like maybe
sending an email to someone who it
shouldn't without my approval. Of
course, there's a risk, but it's like
I've I've weighed up carefully like that
risk compared to these productivity
gains that I'm making. And to me, it's
worth it. And I guess you just have to
do that in your own head. But like
there, of course, there's so many edge
cases with this. It's like you hear like
horror stories about agents doing things
for people. And it's like you hear the
same thing about people getting hacked
on the internet. It's like just cuz
someone got hacked or you've been hacked
before doesn't mean you're not like
going to stop using the internet, you
know?
>> Yeah, definitely edge cases. And I think
like the last line is like data is one
thing. I think like connecting your
something that has like your credit card
and it could like go buy things. I feel
like that is kind of the last line of
defense, but like
>> you could just do like for a bunch of
things like draft only or take it to the
finish line and don't click the button
at the end.
>> Yeah. So I would say like if you're
really like conscious about it like give
it these tools but give it what's called
a readon access. So it can't do actions
it can just read the data. For example,
read emails but it can't send anything.
And that's just a nice way to start out
and then you can like increase the
permissions as you start to like feel
more trustworthy. I give Claude full
access to everything. Stripe everything.
It's it's got full access. I'm
comfortable doing that and it's a risk
I'm willing to take on. I've been in
power using it for like 8 months and
I've never had any issues. Some of those
other harnesses like Open Claw and like
Hermes and stuff, they are built in a
way where the agents like not going to
ask you for permission as much. The way
that like these big companies like
Anthropic and and and OpenAI have built
like Codeex and CL code. The models are
very good at asking you for permission
for stuff. Like pretty much anytime
sometimes it gets annoying for me. It's
like I'm trying to get it to answer
emails, but it'll ask me for permission.
Hey, should I send this? I'm like, yeah,
send it. Like I've given you
>> like allow access, allow access, allow
access all the time. Yeah.
>> Yeah. Yeah. Yeah. Yeah. Exactly. So, I
I've never had any issues personally,
but it's just something you have to
weigh up. I mean, like there's two sides
with the security question. There's like
obviously giving it your tools and then
it going and like taking actions. That's
that's probably my least cons like my my
least concern is it taking actions on my
tools cuz that's like not that big of a
deal. The other security side of things
is like the models and like putting
sensitive data through it. And this is
like a big thing and like a big problem
I'm trying to like solve with working
with some of these large like enterprise
companies is it's like some of them have
like literally said like we don't want
any personal information company IP
running through a clawed model because
they're not convinced that like the claw
stealing their data.
>> Yeah. So that I mean that's a whole
whole another question but for like the
personal level just like getting the
tools connected and getting it to start
doing things for you.
>> Yeah. just start with read only access
if you don't feel comfortable and then
you can escalate it up as you go.
>> Nice. Um let's talk about a couple of
the tools that are like musthaves for
you on your side. I want to share some
of the stuff on our side as well too.
>> 100%. So if we think back to like our
little agent loop of observe, think act.
We've tuned up that observe step with
context and now we're going to tune up
the act step. Right? These are like the
top six things I reckon that everyone
should have connected up to their agent
or claude, whatever you're using. First
one is Appify. So Appify is essentially
a library of online scrapers and these
are like scrapers that people have
built. Let me just go into here.
>> So it's almost like a third party like
sub agent that helps you scraping.
>> Yeah, exactly. So it's like a
marketplace. There's thousands of
different scrapers you can use here. Oh
wow. So you see on Ampify, right, we've
got Instagram profile scraper, YouTube
transcript scraper, Reddit post scraper.
>> Wow. This eliminates like some people
are using like Grock as an LLM under the
notion that it could scrape Twitter/X,
but now you can use Aify.
>> Exactly. We've got LinkedIn post
scraper, Google Map Scraper. There's
like Craigslist scrapers. You can get
like get lead scrapers, email scrapers.
>> These are all individual MCPS.
>> No. Oh, it's just a
>> and can use what the check. [ __ ] I
never heard of this one. Is it? It is.
Everyone needs it. Like literally
there's there's almost nothing that
Claude like can't scrape off the
internet for you with Ampify connected.
>> Damn.
>> So, this is one that no matter what
tools you use, everyone I think needs
this. They've got a pretty generous free
plan. It's like a five Yeah, like $5 a
month worth of usage.
>> Yeah, we're going to throw another zero
too behind that. This is crazy. Holy
[ __ ] So, Ampify, everyone needs to have
that one set up. The next one is Fire
Crawl. So, by default, like pretty much
all these agent harnesses as part of the
axe step, one of the tools they all have
baked in is a web search.
>> But this is super like simple and
primitive. It's like if you were to ask
the agent like, "What's the price of
Bitcoin right now?" It can just web
search and get the price. It's like more
information gathering.
when it comes to like things like, "Hey,
scrape the branding for this website or
like go look at this website and tell me
like why they're or like do a CRO audit
of my website."
>> Less subjective, more subjective, more
like context strategy type stuff. Not
like what's the this is the question.
This is the answer. It's more
>> No, no. I I would actually say it's it's
more that like the web search it can't
look at the website. It can't actually
like Does that make sense? We
>> can't extract raw data almost.
>> Yeah. Yeah. Exactly. Like web search is
just to like find like facts off the
internet.
>> Whereas fire crawl allows it allows it
to actually like read the page not just
search it like look at the page. So it
can extract like it can tell what the
branding colors are. You could say like
do a CRO a order of the page and then it
can use firecrol.
>> That's the first thing my brain went to
from a CRO and merchandising perspective
from an e-com store. That probably is
big.
>> Yeah. Exactly. So, you know, if you if
you said if if you if you didn't give
Claude Firecrol and it just had default
web search and you're like, "Here's my
ecom store. Like, run a CR order." It
can't it can't see the store. It can
just like maybe like search some
information about your website, like
what you sell or like a little about
section.
With with with Firecrol, it can actually
like look at your page, scrape the
entire thing. It can see like what
buttons you have and where, and like,
you know, it could see like, you know,
your buttons are the same color as the
background. like first fix, make the
buttons like brighter. So like stuff
like that, it can give Claude like the
ability to actually like scrape websites
properly and across across the default
web search, fire crawl, and amplifier.
There's like pretty much nothing that
Claude like can't do for me across the
>> And this is just mitigating the back and
forth of just bottlenecks. It's just
getting you more and more and more
context.
>> Yeah. Yep. And it's less it's less um
human in the loop steps, right? Like if
you didn't have ampify or fire
connected, it used ampify before when we
did the scrape my Instagram account
example. Damn.
>> So, like if I didn't have appy
connected, it probably would have said
like I can't connect to Instagram. Like,
can you send me the videos or something?
>> Context for people out there listening
too, like you don't have to say
anything. This is embedded in the brain
on the LLM platform.
>> Yeah. Yeah.
>> You can think of like you can think of
our um really capable employee that
we've just hired. You can think of him
as he's got his desk, right?
>> Yeah. and he's got next to him, he's got
a bookshelf full of skills which we're
going to get to, but he's also like got
the tools like laying around like he
might have like uh, you know, he's got
his computer. He can see like he's
almost got this up, right? If we go
applications,
you can imagine that our ultra capable
employee has this screen up, right? So,
anytime we give him a task, he doesn't
you don't need to tell him go use
Ampify. It's like if you tell him, "Yo,
scrape this brand's Instagram page and
give me the top best performing videos."
>> He's going to just pick the right.
>> Yeah. He's just going to look at his
apps. Okay, cool. Like, I should
probably use Appify. And then he's just
going to go use it. So, you don't need
to actually specify it.
The next one is Composeio. Now, this is
like some absolute source I'm going to
put you on, right? Some quick backstory
about how I found out about this
wonderful tool. I had everything like
baked into Claude. I had all my
connectors built into Claude like I
showed you before and Claude went down
for an hour the other week. Like the
servers were down and I had like so much
work to do. I'm like, "Oh my god, I I
and I and I can't work without Claude
now." So I had to open up like Catgt's
codeex to like work in there. But and
and the way that my folders set up, you
know, we went back to the demo folder
before. I've got like this OS folder
stands for operating system where I work
out of which I'm going to show you guys
after this. So the way that that's built
is I own all the context files. I own
all the skills. They're not embedded in
claude. I can open that folder up in
codeex and keep working, right? That
that part was fine. The issue that I
found was I had like 50 different tools
connected to claude. And then I started
to get to Codex to work and it like
didn't have any of my tools. So I
started to go through and add in like
all these tools and and I was like there
must be a better way. That's how I came
across Composio. So what Composeio does
is we can open this up. You can see
connect apps. They have way more than
than the Claude desktop app. They have
thousands of tools that you can connect.
Same way with Claude, like I showed you.
It's one click login. You click connect,
sign in. They've got everything under
the sun here. And then
>> is this like an MCP for MCPS?
>> Yeah, it's like an M. Yeah,
>> it's like an MCP. It's like a bridge
basically.
>> Yeah. Yeah. So I will connect all my
tools in. And another thing, another
thing is in the Claude desktop app, you
can connect Gmail, right? And log in and
you've got Gmail connected. But if
you've got three Gmail accounts, you
can't connect all three in Claude. Pain
in the ass. And which most people have
like multiple email accounts. So in
here, you can see I've got three active
Gmail connections. That was another
thing that I liked about it.
>> So in short, you're basically saying
that connect all of your MCPs to
Composeio and then if you're in aio
>> and then you and then Composio is like
that single MCP with all the MCPS
underneath it.
>> Damn. So you can see here in in Claude
in connectors Composeio. This one won't
show up as a browse connector. You'll
have to click add custom and then add
the URL which I'll show you how to do
later.
>> And
yeah, we can see I've just got
Composeio. So I could like theoretically
just have that. I don't need the others.
You can even connect the other ones I'm
talking about ampify and firecrawl. You
can connect them through composio.
So
>> damn
>> composio that means like for my setup
trying to like be you know robust and
protect myself against like platforms
going
>> you need to go to your bunker in
Argentina cuz the world is crumbling you
can just you can just hit up composio
and just bring all your [ __ ]
>> the MD files the skills of composio
>> [ __ ] so when I in future now if I want
to work in codeex I don't have to
connect all these tools again I just
have composio and then if I want to try
out let's just say like complexity
computers is another one I haven't
really used it much just another car,
another different flavor of the same
>> But if I decided I wanted to use that, I
could just I'll just compose it to there
one time. Yeah. And then it's
everything's hooked in. So for everybody
out there listening, I want to really
really stop and think like where my
brain used to go and this is the problem
that I used to have with my team. It's
like, okay, I've never been on codeex.
So if Claude went down, I'd have to go
elsewhere. Like I don't even know if
it's as simple as just clicking the
connector to get Composeio, but you
could literally just go into chat GBT
and just say, "How do I connect
Composio?" And I think everyone just
needs to understand that all the
answers,
>> all these in between the lines answers
are literally in the LMS. You just tell
them this is what I need done and they
can do it for you.
>> That's how I learned all this.
>> Like the only like all this stuff I'm
sharing with you today, I just it's
pretty much all learned from asking
Claude.
>> Yeah. People need to go through the
exercise of like if you're not a
technical person and you'd have to bug
somebody on your team to do something
small for you,
>> force yourself to not bug them. And what
may take you or them a minute, two
minutes,
>> it'll take you 5 minutes, 7 minutes, 10
minutes if you're not that technical,
but you will get to the answer. So, it's
more just like
>> figuring this [ __ ] out by yourself is
huge.
>> It's just like, yeah, massive unlock to
get in the habit of just asking Claude
any stuff. It's like, how do I do this?
Like just just ask what it will tell
you.
>> 100%.
>> And if you start following the steps and
get stuck and you just screenshot this
error, paste it and call it. Yo, hit
this. Like what do I do? Like
>> that is a what you just said right there
is a bar on its own. That's a cheat code
on its own. It's just doing that. It's
just screenshot. This is where I'm at.
Can't get past it. What do I do?
>> Let's keep going. So these Chrome Dev
Tools.
>> So the next one's Chrome Dev Tools. So
this is a an MCP built by Google Chrome.
Maybe we can connect that through
Composer. I don't know. I haven't tried.
But essentially what this does is it
allows Claude to spin up its own browser
if that makes sense. So like I use this
a lot if you're building like websites
or apps and Claude like wants to check
its own work cuz you know you might say
like build a website you know these
colors make it look nice the brand
guidelines it might build it and send it
to you and it's just like wrong. It's
not at all like what you wanted. Chrome
Dev Tools is like one I use that allows
Claude to like be able to check its own
work in a sense. Like it can spin up a
browser and look at the website it's
built and go, "Oh no, that's not good.
I'll keep working on that."
>> As opposed to like in platform or a PDF
file. It's like an actual browser.
>> Yeah. It's like an actual browser. Yeah.
And like another thing that I use it for
is I don't know another example the
other day I saw on X someone was putting
together like a register for like
founders in Australia like almost like a
a phone book of of founders in Oz. And I
clicked I was like this is cool. I'll
sign up. And then it was like a, you
know, bunch of questions and it was
going to take me like probably 20
minutes to sit there and answer them
all. So I just like got the link, pasted
it into court. I'm like, yo, can you
apply to this website for me? And I just
kept doing other work. And then because
it has all the context about me, it knew
exactly how to answer all the questions
perfectly and in my voice and but like
without Chrome DevTools, it couldn't
have had a browser to like go and
actually enter it in. Little bit
different to firecore. Firecol is like
more scraping, but this is actually like
it can take actions on a browser. So,
Chrome DevTools is a great one to get
set up. These are pretty much all free,
by the way. Another one is Playriter.
This is one I use. Same vibe as Chrome
DevTools. It's just another browser for
Claude to use. Chrome DevTools spins up
a fresh browser. So, it doesn't have any
of your sites logged in, right? So, if
you ask Claude like go send 100 LinkedIn
connections to this list and it's just
got Chrome DevTools. It'll pop up the
browser and it'll be like LinkedIn login
and it'll just get stuck.
>> Whereas, and you might log in on that
browser. Yeah. But then in a week's time
if you go to do the task it spins up a
new browser again it doesn't have the
login saved. So playrider just allows it
to control your current browser
>> and that like hypothetically like that
could automate like one password with
filling in the passwords advanced fields
for like your address where you want to
ship it to stuff like that
>> on on one password real quick. That's
another one I've connected one password
to claude. So
>> risky business bro.
>> Yeah
>> that's a risky one.
>> Yeah it is.
>> I don't recommend that for people out
there. That's that's like the last
that's the last level.
>> Yeah. What happens though when it's on
my computer and it's like getting a
password, it pops up with the Touch ID
and I have to like approve it with my
finger. So
>> I stand corrected. I like that.
>> Yeah. Yeah. That that's kind of cool. I
feel like Iron Man or something.
>> Someone might come for you and drop off
your finger. [laughter] That's a
different story. Higfield's the chosen
one.
>> Higsfield's the last one. So Higfield is
>> dude, they're they're dominating.
They're crushing the AI content space.
Like they like
>> their playbook is just like I don't
know. It needs to be studied. like they
they've just done an incredible job with
like influencer marketing and like just
they've made themselves at the name as
the top place to go for like AI content
like images and video.
>> Y
>> so you know
>> and that at its simplest form guys is
you just prompt it and it creates images
and videos for you like at its simplest
form.
>> So that's a great tool in itself. Like I
I have some buddies that like live in
there and make some like insane AI
content like UGC videos and static ads
and stuff but by default Claude does not
have a way to generate images. So like
Chad GBT does they've got an image model
whereas Claude doesn't. So this is
especially one that's useful for Claude.
I mean even CHBT doesn't have video
generation whereas Higsfield is
basically like the composio but for
image models. So it basically pulls in
Google's top models, CHBT's top models,
like all the Chinese ones that are
getting cracked as like it pulls them
all in and allows you to just access
them through one MCP with Claude. So to
show you the Higsfield MCP in action,
this is just a demo. I don't I don't
work with Bloom at all. Shout out Greg.
He's
>> shout out Uncle Greg. Great episode. I
don't know if it's going to be it should
probably be the episode before this.
Absolute legend.
>> Yeah, I had the pleasure of meeting him.
He's a great dude.
>> Great.
>> So Bloom Creing Gummies. I just gave it
a product image and it's like yo scrape
competitors ads running in the meta ads
library and generate 20 static image ads
using the Higsfield MCP and then it used
firecrol to scrape the meta ads library
and then it's connected to meta as well.
This is like an official connector that
Met have released so you can get your
agents to like do ads. Another great one
I recommend. But the Higsfield one,
right? You can see it's gone through and
it started like generating static ads
with Higsfield. And like these are these
are pretty good. I mean some of them
it's always going to be the case with
ads like it's you have to generate like
a hundred and there'll be like maybe
five to 10 really good ones. So you can
see here it's like it's done a bunch of
different static ads and like the as a
as an ads background. I would run like
that's a that's a that's a cool one. I
mean the the gummies don't look perfect,
but we could just have added an image of
the actual gummies and then it would
have nailed it perfectly. The like
that's a great ad. The product's a
little squished. That's a that's a
[ __ ] phenomenal ad. That's a great ad
too. Creatine for the girls. Like you
can see here, it's it can actually like
use this MCP to like start creating
static ads. And um this is just off a
really basic prompt, but if I was trying
to like automate as much of the ads
process as possible, I'd build a bunch
of skills around like static ads and
like headline writing and copywriting
and stuff, which we'll dive on soon. But
that's just to show you like the
Higsfield MCP and like some some use
cases.
>> Love it. Guys, this episode is sponsored
by Ketone IQ. I often sit down in this
chair for over three hours and stay
super dialed with ketones. It's a
completely different category of fuel
for your brain. Clean, [music]
sustained, and no crash. I take a shot
before recording, before deepwork, and
any session that I need to be sharp.
Head to ketone.com/open
residency for 30% off your first
subscription and a free gift with your
second shipment. That's [music]
ke.com/openresidency.
Guys, this is something that I use every
single day. Non-negotiable. Give it a
try.
This episode is brought to you by
Momentous. Most Americans today are
walking around deficient in omega-3,
vitamin D, and magnesium. Three of the
most foundational things in energy,
recovery, and sleep. And even if you're
already taking supplements, the industry
is so loosely regulated that you really
don't know what you're taking. I know
the team behind Momentous personally,
and that's why I trust it. Every
product, every batch is third-party
tested [music] for NSF, certified for
sport. So, what's on the label is
exactly what's in the bottle. It's why
over 200 pro sports teams trust them,
and why [music] my whole stack is
theirs. Omega-3, vitamin D, magnesium,
and creatine every single day. If you're
putting real effort into your health,
you want to know that it's actually
working. Go to live momentous.com. Use
code openresidency for up to 35% off
your first order.
All right, guys. All right, so we gave
you guys those six core tools that you
need and then obviously some MCPs that
we both like individually and as a
company. Let's transition into skills.
That's for the newbies out there. Just
to retouch on it, it's basically SOPs
for for AI. Yeah.
>> Why don't you just dive into
the playbook on how to utilize skills?
>> All right, let's do it. So, I would say
skills is probably like the one of the
most important things to master with all
these agents. So where we're at now is
we've got our like really capable new
employee and we've, you know, given him
all the context. He knows everything
about us in the business. Now we've
given him tools so we can do work. He's
already now going to be great. You can
now start giving him work. He's going to
do an amazing job. But we want to be
able to train him on our processes and
the way that we do things. So an example
of skills, they're literally just you
can think of them as SOPs for AI. The
same way if you had a VA that was
helping you with like customer support
for example, you would probably have a a
Google doc built out where it's like
customer support triage workflow. Step
one, log into monday.com. Step two,
navigate to this section in Monday. Step
three, find tickets. Do you know, you
would outline it really really like like
simply to a VA like so it's literally
foolproof so they couldn't screw it up.
We're literally just taking exactly that
but giving it to an agent. And instead
of a Google doc, it's just a markdown
file. And to illustrate like the problem
that skills solve and how useful skills
can be, I'm just going to give a simple
example. So, let's just say you need a
proposal written and you've got your
tools and and and context all set up and
you start a new session and you're like,
"All right, create a proposal for this
new client. Like, here's the price,
here's the offer, blah blah blah blah.
Send it off." it will give you back like
a PDF and like very rare it's going to
oneshot it exactly how you want. You're
probably going to have to go back and
say like I don't like how you put the
price at the top should be at the bottom
instead. Like the logo looks a bit funky
like put it in the top left. Like change
the colors of the text from blue to
black.
>> This is the V1 world of chat. Just back
and forth, back and forth, back and
forth.
>> Yeah, this is still like well no this is
using agents still. Like you've still
got agents like I mean this applies to
chat too but but I'm talking about like
an agent here. Like let's say you're in
code or co-work or getting it to make
this proposal. You have to just explain
your preferences, right? And eventually,
maybe after 15, 20 minutes of back and
forth, you're going to it's going to
deliver this proposal that you're happy
with, right? Great. You had the proposal
written, but what happens in a week's
time when you want to create another
proposal for another client, and you
start a new session, and you're like,
client." Now, it's it's not going to
remember any of those preferences. I
mean, you you could have it save it to
that memory.mmd like we did earlier, but
it's it's not the right place for it.
You're you're you'd be you'd be making a
deacto skill. The the best place to have
that is in a skill.
>> So, if you if you're starting a fresh
chat and said to create the proposal, it
wouldn't like remember any of those
preferences and you're just going to
like not you're going have to go through
that whole process of explaining
everything over again, which is just a
pain in the ass. Whereas with skills,
you would tell Claude, you know, save
this proposal as like a skill. Save this
process that we've just been through as
a skill so I can use it again in the
future. And Claude would build a
proposal skill and it would save
everything that that had just done. It
would save like all your corrections. It
would save like it would probably save
the end result as a PDF into the skill
as well as like a reference for it to
look at. And then it packages that up
nicely into into this like skill folder
and saves it. So now anytime in the
future if you ask it to create you a
proposal, it's just going to like whip
up this SOP that it's written for
itself. Literally a markdown file of
step-by-step guide and then it will just
follow it to a tea. And then you can
basically just set like train your AI
agent on exactly how you like things
done and exact processes.
>> So duplicatable. And I think anybody
that's advanced, I mean, how many skills
do you have? I would imagine an insane
amount. Yeah, we have a lot.
>> Yeah. Yeah, I've got a lot.
>> As far as this goes, like a very very
easy example and we're going to dive
into the different ways in which you can
execute against this. Like
>> we're going to build a live skill.
>> Yeah, we're going to build a live skill
guys. Like even just think like social
media captions, you know, from a
formatting perspective. You can say
hook, body, CTA, tone of voice. You
know, I like to not be, you know, so
professional.
>> Those hashtags out of there.
>> Get the hashtags out of there. And then
an example output you have as well too
is just like yeah these are you know 10
different examples of what it looks
like. And I think like the key thing to
take out of this is is repeatability as
well. Like you execute it and nail it
one time
>> y
>> and then you're done.
>> Yeah. Pretty much. And the way you can
like think of of how our really capable
employee uses skills is you know I was
talking about before he's got his desk
and his computer. You see his apps and
stuff and he's got this bookshelf next
to him of SOPs essentially. and the
anatomy of a skill. Let me actually open
one up. I'm going to pop one open and
show you what it looks like. So, if I go
into like my OS folder, which is like my
folder that I do all my aentic work in,
I have this hidden folder called Claude.
Don't let this scare you. This is just
like literally a Think of this as like
Claude's settings within that project.
We can just open this up and we've got
this folder called skills with all my
skills that live in there.
And I can go into, for example, like
YouTube titles is a skill I've built,
right? And we can actually open this up
as a folder. So, so this is the skill in
its entirety. It's a folder. And we open
that up. And what's inside the folder?
We've got a skill.md file. And this is
literally remember like MD is just a
markdown file. It's literally just a
straight markdown file with my YouTube
title process. And if we look at the
anatomy of this skill, every skill is
kind of broken into three parts. You've
got the name, description, and then
contents. So the name up the top is just
YouTube titles. The description is
generate YouTube title options for a
video topic. Use when the user says
generate titles, YouTube titles or title
options for my video. So you can see
like this is separated from the main
content by those two little lines,
right? And the way you can imagine this
in your head is so you've got this
bookshelf that your agent has a
bookshelf next to it and the books are
all like stacked in, right? And on the
spine of the book is just the name and
description of all the skills. So he's
got like, you know, this full bookshelf.
He would have like the YouTube titles
with that description, maybe YouTube
thumbnails along with my other hundred
skills. And then when we ask Claude or
our agent, yo, can you uh write some
YouTube titles for this video? He's
going to go look at his bookshelf and go
titles. Read the description. Use when
user says titles. That's probably the
one. Pull it out. Then he'll read it.
And then he'll load that con like all
this content in. You got to think about
that example where I was showing like
the the bar chart how it fills up with
context. And it's already like got a
little bit loaded in each session, which
is your your claw. MD and your other
files. And then alongside that, he's got
basically loaded in all the name and
descriptions of every skill that you've
got, which is like probably hundreds.
and that sits in his like taking up a
little bit of that of that context bar
of all times.
>> But it's not like if he if he loaded in
the contents of every hundreds of skills
>> that bar you wouldn't even be able to
like give it a message and it'd be like
100% full. So the way that skills work,
it's called like progressive disclosure
without getting too nerdy. And it's
basically like each of the name and
descriptions just get loaded in at the
start of a new session. And then when
the task like requires like you know
YouTube titles and he sees the YouTube
titles book then it'll open that up and
then it will load the full skill
contents into the session window.
>> Does that make sense?
>> Yeah. And we're going to walk through
that so people can kind of see it in
action.
>> If if we look at this title skill, you
can see that it's literally an SOP like
you would write for a VA. Hey, step one,
look at this. Get the topic. Find the
video topic.
You know, if the user gives it to you,
extract it from the the folder or file
name. Step two, competitor research. Use
Ampify. It's telling it to use that
tool. Remember Ampify
to scrape top performing videos on the
same topic. Target channels. And it's
given some ones that I've I wanted to to
do. what to extract the video titles,
the title length, the structure, use of
numbers, capitalization patterns, and
then it's telling it like how to use
ampify. By by the way, I didn't write
this skill. Claude will always build
skills for you. But you can see like
this is just to illustrate is literally
just a document with step-by-step guides
that you're giving to your employee. And
then step four, generate the title
options, categories, curiosity gap,
benefit first, contrarian number. And
this is the skills is where you package
your IP up into like by default Claude
is not going to write good. If you go a
new session with Claude and you're like
write me YouTube titles, it's going to
suck. It's like the titles are going to
be like so bad, so generic.
>> Yeah,
>> you I know you've tried it.
>> I mean, dude, we're we've tried all of
this stuff, too.
>> And it's like
you package like I've got a way that I
like my titles written that's unique to
me and how I think about titles and I'm
putting that IP into this skill,
teaching it how to think. And the same
way like if you're let's say a media
buyer, if you tell Claude like can you
handle my Facebook ads for me? It's it's
it's not going to know the pro
>> your unique sauce. I mean, yeah,
>> you know, we talked about it before like
>> obviously there's taste and judgment and
creativity that goes into kind of all
these MD files and skills. Like
>> enterprise value is going to live within
like what these skills and MD files are.
That's such a good probing question to
kind of see how deep someone is in in
this space is like what does that look
like? Yeah. Been looking under the hood
on our side. It's [ __ ] out of
control. I can only imagine what yours
is. So, let's just take a step back on
the skills side of like what it actually
looks like to make these skills, whether
you're intentional about it or kind of
your LLM kind of does it for you. What
does that look like?
>> Yeah. So let's let's run through that
for creating skills. In my mind, there's
two ways to go into it. There's goal
first or process first. So the goal
first way is to literally say up front,
build me a brand guidelines skill and
then here's the brand guidelines PDF and
it'll go and build it for you. By the
way, Claude has and pretty much all of
these agent harnesses should come, most
of them do, with a skill creator skill
built in. So it's like skill seion.
You've literally got a skill that has a
step-by-step guide on how to create more
skills. And that's like not one you need
to download. It comes default. I know it
does with Claude and most of the others
as well. So, if you tell Claude like,
"Hey, build me a skill for brand
guidelines," he's going to look at his
bookshelf and see the skill creator
skill and like read it and then know
what to do.
>> Insane. So, in order to make like a
brand guidelines skill, which is a great
one to give your agents so they can have
like a just a repeatable playbook on how
you do branding for your, you know, lead
magnets, PDFs, websites, whatever.
Trying to build that the process first
way doesn't make sense. And I'll just go
straight into like what the process
first way is. So, you've got goal first,
which is, you know, just telling it up
front, build this. The process first way
is the way I build 90% of my skills. And
this will be like I will start a new
session with Claude and I will, for
example, like I'm going to demo my ad
scraping skill. So to build this what I
did was I started a fresh session and
I'm like okay here is the ads library
link for XYZ competitor
please use firecrol to scrape like their
entire like ad library and download it
send. And I went and did it and I was
like all right and now please analyze
every single one of these ads and then
give me a breakdown of like these
metrics the ones I care about send. I
think for people too from like a macro
perspective, uh top down versus bottom
up is basically what this is right now.
It's just how to how to look at it. Is
that
>> Yeah. Yeah. Top down versus bottom up is
a great way to put it's a good way to
put it.
>> And then I'm like, you know, Claude, go
find all the different landing pages
they're running ads to and like send it
off. It went and found like, yo, they're
running it to these five pages. Okay,
use Fire Crawl. I want you to take a
screenshot of all the landing pages and
build a a master doc breaking down each
of the landing pages they're running the
ads to. And then anyway, I went through
this entire process and then I once that
was done and I'm happy with it like I've
taken Claude on this almost like a
journey about how I want this process
done then I will go like well I know I'm
going to want to do this again in the
future. I'll just tell Claude, okay,
create a skill for this process that
we've just done and call it ad scraper.
And then it uses the skill creator skill
and and packages it up for me. And now I
have this skill which I'm going to show
you right now. It's it's absolutely
insane.
>> Blow my mind. Let's do it.
>> Cool. So you can see I've got the ads
analyst skill and I'm going to do a
little demo here. So how this skill
works, I just have to give it an ads
library URL and it will go through like
I said before and scrape all the ads
landing pages and build me like a full
report
>> under the notion that you already have
it connected in your LLM like what we
talked about earlier with the MCPS. It's
it's it's scraping from all the
>> it's using like some of those tools like
ampify and firecraw to scrape them.
>> Cool. So, this is my buddy Ollie. We're
just going to scrape his
>> shout out Uncle Ollie, baby. Let's go.
>> Ollie's Ollie's a gun. He was one of the
ones that got me into into content. He's
an OG of the AI space.
>> Let's go.
>> All right, cool. So, include code. And I
need to make sure I'm using my OS
folder, which is where I pretty much do
all my operations out of because that's
where my skills live. And if we go into
here, we can go slash and it'll bring up
all the skills we have available. Hey,
ads analyst. Ready? Orchestrate
competitive ad research, extract ads,
create a report, and we just click that
ads analyst. Literally, we just give it
the ads library link. Bang. We just send
it off. Go. And now that is going to
start cooking away for us. So, Mark, I
wasn't going to make you wait another 20
minutes for this to to run through. So,
I actually ran the exact same thing
earlier before with the same ads library
URL ads analyst skill. And you can see
it finished up the analysis and it
basically gave us a master report, a
bunch of things, and it deposited it
into this active folder that I have. So
within my OS folder, which is literally
like a folder on my desktop, it's got a
bunch of like subfolders within it where
I kind of just like keep things
organized. And I like to have this
little active folder here where I just
kind of like just miscellaneous tasks
that aren't really related to anything
specific. It just creates a new folder
to work in. So it's you can see here
we've got this ad research one where
it's just sort of deposited it in there
and we can open that up here and it's
built us a few things. It's given us a
master report in markdown format which
we can open and read through. You can
see it looks exactly like a doc you'd
see in click up or notion and we can go
through as well. So and just like
>> wow there's a linked version that links
out to the individual ads.
>> Yeah. So
>> Jesus.
>> And you can see these other reports it's
created. Also, by the way, it's
downloaded all of the assets. So, you
can see here like pretty much every
single video.
>> I wonder if like in the desktop app
it'll pull them up, but if we're going
to show and find it, right, it's
downloaded every single one of his ads.
>> Oh my god.
>> And it has done like a deep dive into
them all. It's pulled out all the
landing pages and it's basically, if we
can open this up, I'll open that up in
Claude in a sec, but it's done like the
full analysis and you can see it's given
us like this final report as HTML
format. So, we can now just shut this up
and I'll open this out so it's a bit
bigger for you to see. So, you can see
that it's given us this full ads
strategy analysis. 54 active ads, 37
video, 14 image, the price point, the
longest running ad, newest funnel 3 days
ago, and you know, it's a competitor of
mine, even though he's a good good
buddy, but same thing in the AI space.
And I can like now go through and scrape
like see everything that he's doing.
It's it's basically talked about like
things to steal or beat. Like what he's
done well. So reframe the category to
justify the price. Like he's selling
like a notion template, but it's not a
template. It's an operating system. So
he's it's it's picked that up and it's
giving that to me is a great takeaway.
>> Name your mechanism. Context
engineering. So the method feels
ownable. You know, longevity.
>> Yeah, I think for context, I mean, this
is a long document. We can spend all day
on it. It's just like you could just
keep optimizing this single skill and
just
>> on the skill side like you can format it
however you like and obviously we're
going to go through the whole entire
gamut but ultimately
>> you can autonomously schedule updates on
competitors like
>> you can do anything you want in relation
to this.
>> I mean look I don't like the way that's
formatted to be honest. So I would
probably go back to Claude and this is a
good point to touch on about these these
skills. It's like the first time you
build them the V1 of the skill is never
going to be perfect. Most of my skills
that I use are on like V5 to 10. So this
one here is like I would then say to
Claude in the chat, I hate how you've
made the landing pages so like big and
full screen. Like I'd like them to be,
you know, smaller and better fitting on
the page. And I would say just update
the skill and it will go and update it
for you, you know, to do anything. And
then next time you run it, skills on V2,
you're not going to have that issue. And
you just keep doing that until the skill
like is eventually sanded down and it's
perfect.
>> Keep iterating.
>> Yeah. Yeah, I mean we can go on and on
about kind of reverse engineering the
skills that you've already built, but I
want to show everybody here and live how
to actually have an idea and build a
skill. We have a mutual connection here
with Beehive. We both love Beehive. So,
why don't we just have you from the
bottom up? Let's just let's just build a
skill in conjunction with Beehive.
>> Yeah. Okay. Great. So, let's let's do a
one of the ideas we had was like a a
subject line skill.
>> Yeah. So, it will basically write
subject lines for my newsletter by
looking in Beehive at what's performed
the best over the last like historically
and then use that to like write the
subject lines. So, first up, we need to
connect Beehive into Claude. So, I'm
just going to show you super quick.
We're just going to go to connectors.
And now, I don't think that the Beehive
MCP is no, it's not native into Claude.
So, what we're going to do is we're just
going to go to plus add custom connector
and then just like close out of these
advanced settings. It's just like
nothing to worry about. And we're going
to name it Beehive.
And then you [clears throat] need the
the server URL here. And like if you
have a tool you want to connect to, you
can just ask Claude like, "Hey, can you
set this up for me or can you help me?"
And it will walk you through it. But in
my case, like I already know what I'm
doing here. So, I'm just going to go
into Beehive into my account. We can go
down to MCP and settings and they've got
this URL for us to use. And like most
softwares that release their own MCP,
they will have a URL exactly like this
for their software.
>> Uh and additional context too guys, BHB
is obviously an email platform and it
was one of the first, if not the first
with an MCP connector from from my
knowledge. So
>> yeah, I've been using MCP. I love it. So
I'm going to show you like we'll grab
this and this URL. So we head back into
Claude and we'll go paste and then add.
And then now it's given Beehive to us as
a connector. And then now let's go into
like a fresh session. I'm working out of
my OS folder, which I'm going to take
you under the hood in there as well in a
sec. And then I'm just going to say,
could you please go into notion and find
my most recent newsletter? And then
I haven't posted it yet. So we need to
create some subject lines. I want you to
then just go into Beehive with the
Beehive MCP and find like all my subject
lines and then pull in like all the data
so you can see which ones perform the
best and use that to help write like 10
new options for the most recent
newsletter. Send that off. And you can
see like with my prompts there's no real
structure or formatting to them. It's
just like plain natural language.
>> Well, that's Whisper. It obviously
breaks it down and kind of
>> streamlines and optimizes the whole
entire thing. Yeah, but I know like some
people are like worried that they don't
know how to prompt it correctly, but
it's like I don't know. I feel like
prompting used to be this thing people
becoming less and less.
>> And guys, anything that you can think of
with this beehive thing, I mean, it
could do cohort analysis. We're going to
do subject lines. I mean, it could look
underneath the hood and see who's on
your list. And if you want to say, you
know, I have this cohort of operators or
athletes and it can extract lists of
those people. Yeah. And
>> do they even have this crazy I don't
know if you're using it like the UTM
tracking where you can also see like on
an individual onetoone basis, this
person with this email has clicked on
these specific emails and like you see
what percent of time they open up. You
could see the source of where they came
from from an attribution perspective. If
>> you're doing like lead magnets, you can
then see like, oh, they came from the
George Heaton lead magnet and then you
could start evaluating like, okay, this
lead magnet was the best lead magnet and
why was it the best lead mag
>> dude, the rabbit hole is absolutely
incredible on this.
>> Yeah, 100%. I've been using the Behive
MCP as well for like so much stuff cuz I
like have my newsletter that I write
each week and it's like a big part of
like what I do and and where I add value
to to my audience and yeah, it's it's
just so good for like subject lines like
data analysis. I was getting my entire
beehive list into resend to send some
like like flows and stuff and it just
like one click moved it in for me. Like
I just said to Claude like create the
recent audience from Beehive and it just
like did it for me with the MCP. It's
just so handy. You can see here like
first kind of in its first few loops
around we've said like go into notion
find my most recent newsletter first and
then cuz it needs subject lines to write
for a newsletter, right? You kind of
have a subject line for you know of an
email. Yep.
>> So it's going to do that first and it's
going through the loop right now to try
and find that email.
>> You can see what it's doing. Hey, while
while that's cooking, I wanted to just
like we can cut this in, but I wanted to
show one other skill I built, but I'm
going to do like a 30 second version, so
it just be super quick. I just want to
blow people's minds here real quick.
>> I don't know if you're ready for this
>> That's it. So, this is a product shoot
skill. I actually built this for a
buddy's company, and we have this skill.
So, that the brand's like Tber Stone.
So, we got like TBA product shoot is the
skill I built. So I would just go
literally I run t product shoot and I
would just get the reference image of
the product. So like this cabinet for
example which can just be like iPhone
images of the product and then I would
run the skill pointed at the image and
then it built it it would generate all
of the production ready images for the
for the thing. So so this little cabinet
you see here that's the reference image.
That's not the final one. Look at this.
That's took that one that you just saw
and now it's built it out into a full
like editorial thing in the same office.
You can see here. Look at that. Ready?
Yeah. It's like a different angle with
someone like touching it. And then it's
got like a closeup of the veining on the
marble
>> and basically the input output is just
giving the the raw like basically CAD
file and then you have a very very
detailed skill that outputs whatever
five or six of these things that could
be like on the PDP page or something.
>> Yeah. And there's another one. This was
the reference image we gave it. And then
these are like what it what it built for
the like editorial. Like just insane,
>> man.
>> Look at these close-up ones. Like it is
so insanely detailed.
>> Yeah. And I I assume that this was
filtered through the Hicksfield MCP or
No.
>> Yeah. Yeah. This was done with Hickfield
MCP to generate the images. And it used
I think like Nano Banana Pro as the
image model, which is Google's one
through Hicksfield though. So that was
like one of my favorite skills I've ever
built. And then we'll cut back to the
beehive example. Okay, cool. Now it's
pulled out like some of the patterns
that win. Like you can see it's got like
don't stack tool names. Your two worst
lines both jammed two tools into the
name. Bold stance. I plus bold stance
wins. I plus boring task doesn't. I want
my don't want my business to be trapped
in Claude. Got a great one. I automated
meta ads in 20 minutes even though I
think that's a good subject line. Didn't
do so well. But you can see it's pulled
the gist of them. It's got like written
out some for this new newsletter that
we've got. And then I can now say as
well and it's see it's got that pattern
as well cuz I usually put this little
emoji at the start of all my subject
lines. So I'm just going to say to to it
now I'm going to say like great can you
please build a uh subject lines skill
that always goes into beehive to check
the best performing latest subject lines
and historic as well and then uses that
data to help write the subject line for
the latest newsletter.
You can see it's cooked away here. It's
created the skill.md for us. And we can
see that it's got the name subject lines
description telling it like what to do
which is going to be on like the little
spine of that book.
>> And that that name the subject lines is
the slash and then it's the the drop
down menu and subject lines comes up.
>> Yeah. Yeah. Exactly. So I'll show you
that in a second. You can see it's
written this doc for itself. Hey, step
one, load the tools.
Step one, pull B performance data.
Step three, find the draft and the hero
story in notion. Then write 10 ranked
options
output structure. Use this shape when
giving them to the reader. It's so
simple. And then we can basically like
that's saved the skill to our claude
folder, which I'll touch on in two sec,
but let's for this example, let's tell
it remember that demo folder that we
built earlier where it had like our
claw.md and stuff in it. Let's get the
skills in there so you guys can see how
this looks when we like continue
building out.
>> And that's adding in like the brand
voice and the stuff like that that we
talked about earlier.
>> Yeah, that project. Yep. Yeah. Yeah. So
I might say like could you please find
the demo folder on my desktop and then I
want this skill to be saved in that demo
folder. Okay. So now it's saved it to
our demo folder. So I I want to just
keep coming back to this example here
because this is like our folder that
we've been building on, right? We
started with the claw.md. We added the
context in the memory in. Now we can see
it's created this little folder for us
called the dotclaw folder.
>> Why is that dark?
>> So the and see it's got the dot in front
of it. Any folders that you have a dot
in front of are hidden. So if we go
commandshift dot on the Mac or period,
it hides it. So by default when I'm
working in my in my normal project, I
don't see those folders unless I run
that little command shift dot command.
It just hides them. I guess just for
like workspace cleanliness.
>> Yep. And then that's kind of like
Claude's settings. And the same way
Claude like when it's loading into a
session in this demo folder, it will
always know to load in claw.md, it will
also always check that. Claude folder as
well. So quite often what I'll do in my
projects is I will put all this inside
the claude folder as well. So that way
like I can just have that hidden and
then I just have in my working folder
that's kind of tucked away and I can
just have my actual files that I care
about. So, if we open this up, it it
created this little folder called
skills. And then it's added in our
subject line skill with our skill.md.
So, that's in there nicely. We can open
that up. And then there we go. Our name,
description, and contents with the
little step-by-step guide for the
subject lines. And then now all like
these context and skills and stuff, I
just like to leave them in this doc
claude folder. And then that's kind of
like hidden away within the project.
>> Yeah. And just to just to tie a bow on
that for, you know, for everybody out
there. So it's basically you have the
skill. It's the SOP on the actual
structure on how you are now going to
build subjects lines moving forward
tapping into the MCP with Beehive for
the most optimized results. And you can
keep iterating and optimizing that
actual skill.
>> And you can change skills too. So like
my I've got a skill called YouTube
publish workflow. And what that does is
if you open up the skill.md it literally
tells it step by step, hey step one, use
the YouTube title skill. Hey, step two,
use the YouTube thumbnail skill. So, you
can see you've got like almost like an
orchestrator skill that will like call
the other skills in sequence.
>> So, he's opening up book one for YouTube
workflow and then it's telling him, "All
right, now you got to get these four
skills and he's like, "Oh, okay." And
then he goes to other books and it'll
keep working. So, like you can
>> create some insane workflows. For people
out there that are worried that your
brain is going to be irrelevant at some
point, I mean, you just touched on
something that is very very relevant is
like
>> the strategic mind is still very very
important to understand the sequencing
of how to look at things
>> because some people want thinking.
>> Yeah. Some people want to look at
thumbnails before they do titles and
without kind of giving it that input, it
would not know that. So exactly
>> the chaining of skills that's very very
interesting and that would just live in
that core skill file and it just would
say check this check this check this
>> just they're just separate ones I'll
show you. So in my cloud folder with my
OS folder I literally just have this
skills here and for example the I mean
that ads analyst one I ran earlier was a
was a orchestrator skill. It called the
ad creative analysis skill. It called
the like creative strategy skill. It
called a bunch of them. And then my
YouTube publish workflow skill. The the
the skills that it's orchestrating are
not within the folder. They're just all
standalone skills. So I can just open
this up and show you this skill. YouTube
publish workflow. Literally says this is
an orchestrator skill. Use these other
four skills in sequence. You can see how
simple that is.
>> Insane. I think we've been looking at a
lot of folders and talking about a lot
of things. I think that there's a lot
for people to wrap their brain around.
I'd love to kind of zoom out and kind of
look at like your machine, your
operating system. Talk about like folder
structure and just see from a macro
perspective like how everything is
organized so then people can see what it
looks like when you put in a [ __ ] ton of
hours and kind of what the end looks
like. So I'd love to just give you the
floor maybe do like a live walkthrough
of what everything looks like on your
set.
>> 100%. So the way that I run things and
this is like the the kind of way I've
landed after months and months of use
and trying lots of different setups.
Guys, real quick. Everything that Remy
is setting up, the loop, the folder
structure, the entire system, he made a
plugin that does it for you, built
around your life and your business. It
interviews you, then connects your
email, your calendar, your docs,
everything. I am not technical. I did it
myself. 90 minutes, one sitting, done.
Sign up for the newsletter below and get
50% off at launch. Enjoy the rest of the
episode. I originally used to like kind
of create these agents. Like I would
have like really granular niche agents
like id have like a you know I mean at
one point I might have even had like a
subject lines agent.
>> And I mean when I say agent you know a
folder like I would definitely have like
a you know marketing like ads folder and
then I would kind of just like open up
those folders to work in if that makes
sense. If I was doing ads I would open
up my ads folder include code and like
scope it to that folder and then work
out of there. But the more I I build and
work, it's like it's just such an
annoying, inefficient way to have
things. I just run everything out of
this OS folder. And some reason as
humans, we love to like give these
agents like treat them like humans, you
know, give them human names and like
give them like roles and things.
>> Marty, we got we got Marty.
>> Marty. Yeah, exactly. Like people people
for some reason they they want to have
like a head of marketing agent, all this
stuff. And it's also what gets lots of
clicks and views on Instagram, you know?
People go like, "I've got my team of
agents doing this." And it's like
really, they're just folders. And the
more folders you have and more
specialized, the more like different
stuff there is to worry about. And the
models are just so smart now that I just
run everything out of the OS folder. And
then I build really really good skills
to teach it the processes. And this is
my OS folder here, right? So I have all
of my like this is all my stuff for my
like, you know, education, consulting.
This is stuff for for comfort work I'm
doing there. And this is just another
like thing just for some notes and
stuff. And
>> so so let's take a step back. So if
somebody's looking at this and they want
to look at their folder structure, let's
just say for example for me like I would
have iconic and I would have open
residency. Like these are the big
pillars in which I want to work in.
>> 100%. Yeah. So you got exactly right. So
you can see like my my folder structure
is literally like I've got like my
overarching OS folder. OS just stands
for operating system. And then I have,
you know, any like big any businesses or
pillars. So yeah, for you iconic and
open residency. You might even have
another one for like you might have a
folder just called personal for like
your health stuff that you wanted to do.
>> And I've got another one that I like to
have. I just have a folder called active
which is just like oneoff projects and
experiments that don't really fit into
like one of the pillars.
>> And if potentially it becomes something
substantial, maybe you'd move it into
one of those other
>> core folders. So for like I created a
folder for this podcast to prep and I
just added it's just sitting in the
active folder. It's loose. It's just
like a separate project that I'm working
on. And then within these like
overarching pillars then I'll have like
the folders within like content,
newsletter, website. So for example in
my real OS folder you can see here AOS
bromy we go in and I've got like content
finance legal newsletter website
marketing and I used to
open claude so I would in the past have
let's just say I wanted to work on
marketing what I would have done is I
would have opened up the marketing
folder and then started working in there
for all my marketing stuff
>> and what's the right way to do it
>> I mean there is no right or wrong way
but I I the reason I was doing that and
my thought process was then it's like
we're scoping the agent to just
marketing work but they're so like the
agents are so smart the models are so
good and their context windows which is
like the bar chart like how much you can
fit in it is just getting bigger and
bigger that I just run everything out of
my OS folder
>> that's under the notion that the OS is
very macro and it's going to pick up all
the stuff underneath it as opposed to
going granular
>> and what I would do right so let's just
say like for my new like a newsletter or
marketing let's say we're creating a new
batch of ads I would sort of have OS
folder open working. And then I like to
just go like this. I would tag the
folder. You can tag folders. And then I
would go, you know, AI with Remy
slash
marketing. See, it's got the marketing
folder.
>> I mean, there has to be some muscle
memory in there and knowing what your
kind of, you know, folder structure is.
I would love for you to walk through
right there just the ads, the slashes,
like what are all those kind of like
quick shortcuts that people can do to
tap in there?
I mean it's going to be slightly
different depending on what harness
you're in, but most of them like codeex
cloud code pretty much all of them use a
slash command and an at symbol type
system. So slash command just brings up
all your skills and like commands
available to run. As you can see, we've
got a lot and but it also just brings up
like settings that are baked into
claude. Like for example, if I want to
change the model, the LLM, the brain
behind this loop, I can click do /mod
and then it's going to bring it up
there. And for context too for people, I
just want to double click on everything
to give people 100% clarity on stuff
too. If you're going to slash skill, is
that like tag teaming it in? What
happens if you don't tag team it tag
team it in? Is it going to automatically
tap into the skill with its own brain or
do you have to be very intentional in
making sure that you're tapping in the
skill?
>> Sometimes I'm intentional. If I say,
"Write me a a hook for my Instagram
video," and I've got a hook skill, it's
just it's going to pick it up.
>> Yeah. But like for example the ads
analyst one if I just dumped in the ads
library link and ran it would be like
what the [ __ ] do I do with this?
>> So naming conventions are important but
obviously like really having muscle
memory in what your folder structure is
so you know when to tap into certain
it's basically just like you have 50
employees you should know the names of
your employees.
>> Yeah. Yeah. Yeah. Yeah. Exactly. Same
thing. I mean like I've got a pretty big
setup and it's not that hard to like
work it out. Like I mean even so like
slash commands is for your skills and
like other little setting stuff mainly
skills but the ats I find is like when
you want to tag files or or folders so
like or documents you know for example
um
>> and should that be under the notion that
if you tag in a folder it is going to
pick up all of the files within that
>> Yeah. Yeah. So you know I I wouldn't
even really have to remember I think if
I just go at marketing right I want to
do marketing it brings up straight away
my marketing folder. Hey, with Remy
Marketing, I can select that. And now if
I give it my task for marketing like
write me 10 new ads for my upcoming
workshop.
Then you will see what it will do is it
will then start to like use that folder
to work in but without us having to
actually open that folder up to work in.
Does that make sense?
>> Yeah. You're you're basically uploading
the OS folder, which is macro, and it's
getting everything under it, but then
you're tag teaming in the marketing
folder to take like an extra good look
at the marketing folder.
>> Yeah, that's exactly
>> I don't know if I explained that good
for you guys out there, but I think it's
uh I'm I'm trying to make this at a
fifth grade level so people could really
understand it. It's just like, hey, go
check this folder. It's probably where
it is.
>> I mean, that's like that's that's what I
meant about demystifying this all. It's
like you hear like people make Instagram
videos and it's like my like my
marketing agent just did XYZ while I was
asleep overnight just worked all night
and it's like if if you don't understand
this you're like oh like wow this is
crazy but in reality they've I I
guarantee you they have a folder called
like marketing and they have some skills
in there that are marketing related and
they're running those skills on some
kind of schedule. So it's like hey every
night do a report on Facebook ads and
email it to me so I wake up to a report.
So so simple. It's just folder skills,
schedule tasks, and then but people like
take that and I don't know whether it's
an ego thing cuz they they like to sound
smart and sound cool or whether they're
pushing some course and that they want
you to think they're like some AI god,
but in reality it's like just as simple
as a skill and a folder and a schedule
task. But people will say, I've got like
50 agents working for me overnight.
>> And that could be anything. It could
literally be clean the inbox. You can
say draft draft email responses,
>> send cold emails or like send LinkedIn
connections, you know,
>> suggest emails that you want to
unsubscribe to if you haven't opened up
the last three, four times.
>> Yeah. So, it's like, you know, now if I
set this up and click create, now I've
got an agent that is managing my email
inbox for me.
>> Insane.
>> It's it's it's a lot simpler than you
think. The last sort of concept that
goes really handinhand with this sort of
folder structure I'm outlining is you
have context tools and skills you can
have them at a global versus a project
level. The easiest way to think of
global is like a holding company and
then the project level is like the sub
companies which is kind of like how I
have my setup and you know things at a
at a OS level like a global or a holding
company level are things that apply to
like every single company. So, if you're
looking at this holding company example,
you might have like context, tools, and
skills that apply to like all of your
companies. And then you might have some
things that only apply to really really
specific companies. So, I have like
skills that live at a global level,
which I can access anywhere. And I also
have like skills that exist just within
a folder. Like remember that one I
showed you before with the product
shoot, TA product shoot? that skill I've
got kept in the folder for the tea stuff
cuz like I'm never going to use that
skill outside of that specific folder.
>> It's so crazy cuz this is really just
like this whole thing. It's a personal
OS but it's really just a holding
company. The whole entire infrastructure
of the folders is like a holding
company.
>> Yeah. Yeah. Exactly. That's that's
exactly the way to think about it. So
another example of like global versus
project or holding company versus
company specific is our claw.md. You
remember that?
>> Yeah. So, I have one that I have in my
OS folder, my global claw.md that I want
to apply to like every single session
I'm in, but I have a claw.md in some of
these sub projects. Like, I'll give you
an example right here. I'll open up the
folders. So, I'm in my OS folder right
now, like my real one, and you can see
that my like, you know, in my I guess
you call like your settings folder. I
have got my my claw. Mmd, my global one
for my OS folder. And this one here
contains like overall preferences about
myself, you know, explains like who I am
as a non-technical founder learning.
That's not going to change no matter
what project I work in. I'm always going
to be the nontechnical founder who's
learning. So, I want that applied
everywhere. I've got in here, these are
loads of lessons that I've built up over
months of use.
>> I've got I saw in there like core
principles, core values. That's a great
>> This is one I have as well. Boil the
ocean. The marginal cost of completeness
is near zero. basically just tells like
the agent the standard isn't good
enough. It's holy [ __ ] that's done.
Like just little things like that,
right? And I want that applying to every
session. But for example, when I was
prepping for the podcast, I created this
little podcast folder called
>> podcasts
>> podcasts, right? And then in here, this
is where I was just like workshopping
and planning everything, right? You can
see a bunch of files. This is just like
a big like work workspace essentially.
But you can see within the podcast
folder, I've got a claude. MD. And if we
open that up, that just contains like
the Northstar context for the podcasts
>> Yeah. So it's it's it's going to be
filtered through the macro cloud folder
for like your principles and how you
think. And then this is more like rubric
and outline. Yeah.
>> So it's simple, right? This workspace is
Rey's base for podcast appearances. What
this folder is for prep, demos,
diagrams, post episode analysis.
That's literally all it is. And then I
just have some other folders in here
where I was working on stuff. But now
when I'm opening up Claude to work, I am
still working out of the OS folder and I
would tag podcasts. Let's say I wanted
to prep for this pod. I would tag
podcasts and then I would give it my
task and it be like what's the outline
for the open residency podcast.
Even if I didn't tag podcast the folder,
it would still go and find it. But I
just like to do it for nice like
cleanliness and peace of mind just to
make sure. And you can see here it's
opening it up. And
>> just to to rebutton down for everybody
out there listening, you brought in the
OS folder. So it's getting the macro
claw and all the stuff there. And now
you're tapping in the skill of podcast
and it's picking up the claw MD file
within there and anything and everything
y around there.
>> And and like now it's loaded in the
claw. MD from the podcast folder. So
you're kind of like stacking them on top
of each other.
>> Yep. And you can see he's whipped up our
our episode plan here, which is which is
super cool. But that's just giving you
an example of like global versus project
level or holding company versus company
level CLMD's skills and tools.
>> Amazing. Amazing context there. I'd love
to just shoot for my own selfish reasons
and I think for a lot of people out
there, they probably are interested in
this as well. just like something that
I've been working on and just how would
you kind of construct it on a a global
versus project level is like me I'm like
on a health journey let's just say I
have my DEXA scan my blood work maybe
I'm tracking my macros this whole entire
flywheel of all these kind of touch
points maybe I have my aura sleep score
which has an MCP and can bring it in
maybe eight sleep
>> and and can bring it in like how would
you architect that global verse project
and then like just give me your your
thoughts on that
>> yeah so I would have my I guess holding
company
OS folder. Yep.
>> And then, you know, you might have you'd
have a folder within that for like open
residency, another one for iconic. And
then I'd probably have another folder
because it's a big pillar in your life
like personal.
>> And then within personal, I'd probably
have a folder called like health. And
then I would just in that health folder,
I would then just dump in all that
stuff, everything. Just dump it in like
Dexter scans, whatever. Just just dump
it in there. And then I would just
create a claw.md just for the health
folder. And I would basically just say
something along the lines of this is my
health workspace where I'm going to be,
you know, analyzing blood work and doing
this and doing that. You might even then
give it like your health goals and put
it in the claw. MD just any context that
your agent might need when it's like
when you want to work on health stuff.
>> And then that's literally it.
>> Yeah. I think like again to kind of take
it a step further, something that I'm
working on now is just like, okay, you
have your separate vendors. Maybe it
takes X amount of time to book an
appointment to go get your blood work,
go get your DEXA scan. You know, maybe
this is your preferred vendor to get the
DEXA scan. You have your Gmail linked
up. Yeah,
>> it knows every seven weeks to email the
DEXA scan place to set up a DEXA scan so
you stay on target and keep updating it.
So, yeah,
>> there's a lot of ways that's something
I'm working on. I'm super super excited
about to get the reminders and automate
the the different appointments.
>> Yeah, I know a few guys that have like
done crazy stuff with like health and
blood work with claw code and and
agents.
>> for me it's simple because it's just
like I saw my blood work that like I
have to work on saturated fats and
sugars and like
>> I just now work through the lens of what
food do I eat and is it bad and
obviously calories and protein etc. but
like what does it look like on those
kind of two realms.
>> amazing. definitely will be talking
about that kind of health whole entire
flywheel that I'm setting up. I think a
lot of people would be interested in
that. What is like the honest kind of
like monthly cost of doing all of this?
Obviously, you talked about, you know,
the $100, maybe $200 a month if you want
to get the bigger plan on your core LLM.
How much are you spending a month on
kind of everything?
everything all up about a
I mean about
a thousand a month but that's across
like loads of tools and you know I'm
it's my job to be on top of this. I'm
I'm testing stuff. I got subscriptions I
forgot about from months ago. Like I'm
Yeah, I know my bills about a thousand
bucks for AI stuff.
>> You need to set up a skill to cancel the
subscriptions that you're not using.
Bro, come on. Come on, man. Come on,
dude.
>> Killing a schedule. I should have I
should have one once a month that goes
to my email receipts and and notifies
me. I should
>> the email unsubscribe thing is is a game
changer. Yeah. So for people out there
listening so
>> $12,000 for basically a year for I don't
know 100 employees.
>> I mean I think it's it's it's a little
bit interesting because all of these big
AI companies right now like OpenAI,
Anthropic, they're all like none of them
are profitable. I mean maybe Anthropic
just got profitable. I have no idea.
>> You have to confirm. None of them are
profitable. I don't know if you knew
that that they are all running at
massive losses because you know the they
on these plans you're buying like the
free plan, the pro plan, the the $30
$100 a month plan, they're heavily
subsidizing the real compute cost of of
like the AI usage, right? So, I think on
a $300 a month Claude plan, you are
actually getting about $6,000 worth of
token usage, but they are subsidizing it
so heavily because they're all trying to
get users. They're just trying to
capture mass market share. And Anthropic
are like really now making a push to IPO
and become profitable. They're starting
to like they've kept their prices the
same, but your usage is going down and
down and down, getting squeezed. They've
just released their latest model. I know
like maybe if you're watching this it's
something new but they've just released
Fable 5 and they've essentially said
like this is our new most powerful model
>> it's not not included in your plan. You
have to pay the real cost of it which is
>> disgustingly high.
>> That's a great that's a great uh new rev
for them to print. I'm sure they're
charging some sort of markup on it but
that's like
>> for anybody that really wants the best
model they're going to pay for it. I
feel like people are going to run that
up.
>> Yeah. So, it'd be interesting to see
what happens over the next sort of year
or two as these companies like they
can't keep running at a loss forever.
They're going to have to start to bring
the costs up. So, it'll be interesting
to see what happens. I think there's
going to be a massive market for like
open source models, basically models you
can run yourself. You can buy like, you
know, couple Mac minis and and run your
own AI models on there. I think there'll
there'll always be options for like
lowerc cost AI. There always will be,
but it just be interesting to see these
big companies like what happens with the
price of like your your plans.
>> I completely agree. We've only talked,
you know, I mentioned the Kia. I I think
that there's only Lamborghinis and Range
Rovers right now. Eventually, there's
going to be some Kia and Hondas and all
that is going to come.
>> There's a few.
>> There's going to be a lot more options I
feel like on the lower end. I don't know
how they're going to figure it out, but
definitely will be. Before we move on to
the next thing from a a folder structure
perspective and you know global versus
project anything else big that you want
to touch on there or do you want to move
on?
>> That is pretty much like all you need to
know and you know I wouldn't get like ke
up about having like the perfect folder
structure. It grows with you over time.
Just start out with with create a folder
called OS. That's that's your first step
and then just you know create your CLM
day add the context in and
those subfolders and like you know
holding like companies within there
that'll just come over time as you start
to do real work like I wouldn't go out
and like try and get the perfect
structure. I wouldn't let that hold you
up from just starting and and getting
work done with these agents.
>> Just start the two layers of AI and
business. I'd love to kind of walk
through your thesis on the different
layers and how you look at it
specifically from a business
perspective.
>> So the way I see it right now, you
almost have like two layers of AI within
within the business, right? It's like
you have your personal operating system,
which is pretty much everything that
we've covered today. This in my mind is
like an extension of you. It's like
you're going to become like an AI
augmented founder. Like the the the way
I think of it is like imagine a dude
with like a robot arm. It's like an
extension of of him and now he can like
climb walls with his arm and do all this
crazy stuff. And that's like this is
your personal operating system, personal
AIOS is kind of where you live in every
day. This is like this claude code thing
I'm setting up with these folders where
I now complete my entire workday from
within there. And you know that kind of
handles all my tasks and my grunt work.
But if you think big picture, let's just
say you wanted to have an agent that
managed your customer service or
customer support, the it would be the
wrong place to have that on your
computer as the founder or you know your
customer support team. Like it probably
wouldn't be make the most sense for them
to have that on their computers. you
kind of like that that's where this
second layer comes in which I call it
more like your AI business
infrastructure and that's like
that's where some of those other agent
frameworks like Hermes and Open Claw
really come into play
>> more advanced frameworks
>> more advanced frameworks and they're the
kind of things like that's why everyone
was running Mac minis are flying off the
shelves because people just wanted like
a computer that was not theirs almost
like a company Mac Mini to set some of
these agents up to run on and but if you
if you look under the hood at these
agents so right now Like I just remote
accessed into my Mac Mini that I have at
home. And you can see here I've got like
a a Hermes agent set up and this is like
the file system you're looking at right
now is not on my laptop. This is the
files on my Mac Mini at home. And you
can see
>> this is remote. This is like the cloud.
>> So just like I have the OS folder that I
work out of, the the Hermes agent is
literally just a folder as well. See
Hermes?
>> We open that up.
>> Yep. It's just like a bunch of folders
and markdown files, you know, like a
soul.md which tells it how to behave
like all the same principles that we
have just gone through today. So when
you eventually decide you want to like
graduate to start using like something
like an openclaw or a Hermes agent for
different stuff to if you want to
configure these agents properly, still
the same thing context tool skills,
nothing changes. It's just like they're
just slightly more advanced frameworks
that you that work better to run like
more autonomously on another machine if
that makes sense.
>> Yeah, I feel like there's a huge
opportunity for I think first steps is
to become acclimated and understand the
tools and do your own kind of personal
OS. But it begs to ask like you know the
question that everybody's asking next is
like okay who owns those files? That's
for the individual person. Is it for the
person? Is it for the role in the
company? And then is there like a
separate infrastructure for a company?
Like how is there, you know, someone
works for a company, they're setting up
all of these skills. Then what happens
if they leave the company? Where do the
skills go?
>> Yeah, it it's like because this is so
new. It's so messy right now. Like it's
really this is all really simple at a
personal level, right? Like you build
skills, you connect your tools, you have
your folders, and they just live on your
computer where you work. It's simple.
But yeah, it it it does start to get
messy when you for example like trying
to work with a team and you know someone
builds like a stock forecasting skill
and then now you know another team
member wants to use it but it's on their
computer and it's like well how do you
give it your team member access? Oh well
you could airdrop them the file but then
you've got two separate copies and
they're going to get out of sync fast if
people are like making updates to them.
So to answer your question like that
there's there's a few ways you can kind
of bootleg it with like tools like
GitHub and and Google Drive and stuff,
but there's no like real clean simple
way that I've found yet to to
effectively like manage skills across
like team members and companies.
>> There's a massive opportunity for
somebody out there to create the
technology that has that connective
tissue where it kind of auto updates and
it can share. For sure. There are some
startups I've seen that are now starting
to solve this with like shared skill
management and there's a there are a few
coming out now and I need to just sort
of get my hands a bit dirty in that area
of things but uh yeah in terms of like
skill ownership like skills like assets
for your company so it's it's a whole
another conversation as to whether like
I think every employee needs to be the
ones building them because you know they
do the task the best like they need to
build the skills but then it's like if
they're assets for the company are you
going to just let them walk away with
the skills bills when they leave or do
you as a company own them? It's a
difficult one. I can't say.
>> Yeah. I feel like at a macro level like
the you know the clawed MD file like
that should be macro and uniform the
stuff that doesn't change that we talked
about like that should be kind of across
the whole entire company as like the
core input in whatever folder they're
working in for the company. Yeah. But
then once you start kind of going
downstream on the skills I mean that's
where like creativity and taste and like
how you prefer things to be done. like
some people, you know, want to work in X
program, other people want to work in in
this program. Some people, you know,
need it in copy. Some people need it,
you know, are a visual person. So, I
feel like that's
>> it is very tailored toward the
individual human.
>> Yeah, agreed. But, you know, I wouldn't
get too ke if you're just starting, you
know, just start just like build your
own skills first and then worry about
all the team management and other stuff
like later. It's just cross that bridge
when you're at it.
>> Yeah. If you're out there listening, if
you do even 10% of what we talked about
today, you're in the 0.0000
1% of people that have graduated from
chat uh everything else
>> past that. So
>> yeah, I think uh that's for another day
that kind of that layer two with the
open claw and and the themes and such.
>> I mean, one other thing I wanted to
touch on too is like every company right
now needs to like push to become AI
native. Like every company needs to like
put AI at the forefront of like all
their operations or they're going to
lose to a company that does. It's like
simple as that. And like I've kind of
been trying to solve some of these
problems for companies I work with about
the best way to go about doing this. And
there's two ways to go about it. And I
and I and I think that one way is more
correct than the other. So the first way
is like you go top down, right? like you
bring in some AI specialists into the
company to like start building stuff and
you know they like they can go and
interview each department and sit with
the employees and like try and build
skills on their behalf or try and like
build agents to take roles but it's like
the employees the people that do those
tasks they know it so well in and out
and there's so much nuance to every task
and if you wanted to bring AI
specialists in to start automating this
stuff they would need to understand so
deeply every task that everyone in the
entire company does they can never do
that. So you either need a lot of people
you need almost need like one per one AI
specialist for each employee if you're
trying to go the top down route where
you build it and hand it to them and say
here use it. I think the best way to get
your company AI native is to get your
employees trained up to be AI literate
get them building their own personal
operating system so then you can the
goal is not to like fire your entire
team. You want to keep your best people,
train them so they can become like 100x
employees. And then you can go on hiring
freeze and continue to grow without
scaling a headcount. And the way you can
start to automate it is you instead of
from the top down, you automate it from
the bottom up. So you train your
employees to become AIT. You get them to
automate their own job essentially. And
then you eventually like have this like
you if you can think of like a a graph
where it's getting like saturated, you
know, the AI saturated
>> infiltrates the whole entire company.
Everyone's
>> Exactly. That's probably like the best
way to go about like trying to get your
company Aative and there's all sorts of
things there that I think work well like
putting incentives around it. Like you
can do like build weeks and
competitions, you know, like running
>> good idea
>> things like you know best like skill of
the month. Maybe you get your employees
to start building skills and then it's
like the best one wins like a cash prize
$1,000. Like I think companies need to
like just make AI like part of their
culture. they need to have an AI call
each week where people like sharing
about what they've built and like really
start to like bring that in as part of
the company culture.
>> I think it's a double-edged sword cuz
some people, you know, the more they
adopt, they feel like that they're
digging their own grave to be fired. And
that lens, you know, you have to talk
about like what are actually skills that
are going to do well in the new economy.
I would disagree though on going two
only. I would probably go a mix of path
one and two. Like for me with iconic
like I had like a a six-page document
that it was let's just say 60% of it was
uniform in nature of like asking all the
appropriate questions like what do you
you know what tools do you use? What are
the processes that you already have?
What are your bottlenecks?
>> And then you're right that other 40% was
100% tailored to that specific
>> employee. Yeah. But with that being
said, it's very tough like someone
that's coming from the outside, they
have to be
>> very very versed in business to be able
to pick up context fast or else it's
going to take a very very long time. But
I think
>> that's the first thing that every
operator out there should do should be
like the download of just like
>> what AI tools do you use? And right away
you're going to get a pulse check on
like where everybody's at. Yeah. And
like just seeing that for my whole
entire team, like that was a big shot in
the arm to be like, "Wow, we got to like
[ __ ] hurry up on this [ __ ] because
with one, you will go faster. If you do
just two, it's going to be slow." I
think it's a mix between one and two.
>> Yeah, I would agree with that. I would
agree with that 100%.
>> All right, this has been an amazing
episode. Computers are down. Let's get
into quick strike. Random questions.
We're just going to rip off.
>> Give me what you're thinking. So, first
skill everyone should build.
>> Email inbox triage skill. Run that bad
boy on a schedule three times a day.
>> Clean your brain.
>> Sorted.
>> Too much noise. Biggest beginner mistake
you see in AI.
>> People being too perfectionist and too
scared to start cuz they like don't want
to screw something up. Like one of my
buddies will shout if you're listening.
They're like we want to use claw but
like we're scared to even like type a
prompt and we don't want to screw it up.
And it's like you know cuz people kind
of think see it as a black box. I don't
know what's going on under the hood. But
it's like just the biggest being mistake
is just being too kind of scared to
start so you don't like screw things up
and trying to be too perfectionist about
the whole thing.
>> Ladies and gentlemen, if you had made it
this far and you don't take action, you
are a loser. One thing you automated
this week most people would never think
to.
>> My laundry.
>> That's that's kind of crazy. Most
overrated AI tool right now.
>> I'd say any AI app builders. lovable
vault any of those and you just just do
it in Claude like just create a new I
say like just create a new project
folder with your holding company call it
website or app or whatever and then just
start getting called to build it out in
there it's like
>> is not going to be signing with lovable
anytime soon most underrated AI tool
right now
>> composio how long until a total beginner
feels the productivity gain like
realistically
>> a few days a few is
>> the task you'll never let an agent
touch?
>> Thinking
and phone sales. [ __ ] AI voice sales
>> One MCP connection that changed the game
for you.
>> I'd say connecting Ampify with all the
thousands of online scrapers. You can
pretty much get anything.
>> Finish the sentence. In two years, every
founder will have a
personal AI operating system.
What is Obsidian and is it still
relevant today?
>> Obsidian is just a simple note-taking
app that uses markdown which is the love
language of AI. The agents love it as we
covered earlier and a lot of people kind
of build out their like I guess second
brains or like that you know context
folder that we built earlier. people
just build like essentially a a
version of that context folder but with
hundreds or thousands of markdown files
and they just use obsidian as kind of
like the the layer to manage that I
guess and I just think um I don't think
it has its place. It's really good for
like solo operators but as soon as
you're trying to like have a setup where
you want your company company and team
members to access
>> too
>> those files too it's messy. I don't
think it's I don't think it's needed
especially for beginners.
What is one skill that's changed the
game for you?
>> The Claude council skill. So, I found
this by a guy called Ollie Leman on
Twitter. And it's actually like my most
viral video on Instagram talking about
that skill. And you know, if you ask
Claude for advice on something, if you
say like, "Hey, why is this a good
idea?" It's going to give you like a
perfectly structured case as to why it
is. And if you say like, "Why is this a
bad idea?" It'll give you the same kind
of case for why it's bad. And you know,
why you can't trust Claude with
decision-m. So this skill essentially
spins up a bunch of different clawed
sessions all given their own persona
like a contrarian that's going to hunt
for what goes wrong like an optimist
five different complete personas to
basically like give their thoughts on
your question and then it gives it
shuffles up their answers anonymously
gives it to five fresh sessions that are
like the peer reviewing agents to like
give their thoughts on those answers and
then finally it gives it into one new
session which is like the chairman that
like looks at all the final results and
gives you the final verdict. I use it a
lot for like decision making cuz it's
just interesting to get like an extra
perspective on it. I'll use it all the
time but like hey should we add this
feature or is it too soon or it's like
what should I work on next on the
project?
>> Sounds amazing. Before I get into the
last couple questions that I ask
everybody I'd love to just kind of recap
and end with like where should someone
out there just start?
>> Yeah. So think of this entire system as
like something it it it grows with you
and it will continue to grow and
compound. And you just need to start by
planting the seed. Just get your
computer right now, create that folder,
the holding company folder OS,
start getting your context files built.
Just open a fresh claw or chat GBT
session and just say like help me build,
you know, a bunch of markdown context
files. We need like an about me, an
ideal customer profile, a product,
services information, uh you know, brand
voice, and then interview me to extract
all the answers from my brain and and
build these documents. And then you add
them into the folder
and start. Here's the big thing. Connect
your tools as well. That's actually
could even be the first step. Cool.
Tools and context you can get done in a
few hours. And then from there, your
goal is going to be to complete an
entire workday from with inside Claude,
whatever harness you're going to use.
That's your first goal. And you know, it
might be a little bit messy at the
start, but you're going to then start to
build skills. It's like if as soon as
you go through a process with Claude
like we did with the subject line skill,
if you ask Claude to take a task off
your plate and you know that you're
going to have to ask that him to do that
task again in in the near future or at
least one or two more times, tell it to
make a skill and then that's the place
to start. You just start growing this
system and you can only do that through
like working with it. So get your your
context files built, get that folder set
up and your tools connected and just
start completing a workday from with
inside your your agent harness. He's
challenged you. Let's do it, guys. Last
three questions. Favorite book or
podcast and why?
>> My favorite book is The Alchemist. I
don't know if you've read it.
>> Polo Quello. I'm Brazilian, bro. Of
course I do.
>> Dude, I love that book. I've read it
like four times.
>> That's interesting.
>> What about podcast? You a podcast guy?
>> Yeah, I'm a podcast guy. I like the My
First Million podcast.
>> Good guys.
>> And the Open Residenc.
>> They're guns, dude. They inspired me to
start doing the newsletter. Like they
they both crushed it with newsletters.
>> Yeah. On the newsletter side. They're
with Beehive as well, too.
>> Yeah. Yeah. They they they've both
exited their newsletters for like tens
of millions, right?
>> Yep. Sam on the hustle and then Sean on
Milk Road, the crypto one.
>> That was a quick one, too. Entrepreneur
or brand that you want to give flowers
to and why?
>> Brand
Composio. Dude, they're so good. Love
Composio.
>> I had never heard of Composeio until
today. We most definitely are going to
put that on steroids and put that on
everything. So, I'm with you on that.
What about a creator you want to give
flowers to and why?
>> Good question. Do you know Ask Cat GPT?
>> she's a beast, dude. Like inspired me to
start creating content and yeah, she's a
she's a beast at content and AI. Big
inspiration for me.
>> Love it, man. I hope everybody enjoyed
this episode for the first time ever in
Open Residency podcast history. I'm
going to ask the Pay It Forward, guys.
Share this with someone that you care
about. Get people started in this game.
This is definitely the most thorough.
but we took a long time to really kind
of orchestrate and outline what we're
going to talk about. So, I hope you guys
enjoyed it. And u we are going to put a
[ __ ] ton of links and PDFs and tons of
stuff down below in the description. And
this definitely won't be the last time
that we do stuff with Remy. So, I
appreciate all you guys. I appreciate
you, man. It's been an amazing day.
Thank you for having me on. I've I've
enjoyed it. I think like this is just
the crash course that everyone needs.
Like if if you're starting from zero,
like this is just all you need, just
this episode and I think you're set. So,
I'm glad we got it.
>> Amazing. Thank you guys so much for
watching. If you got value out of this
episode, every [music] single principle,
framework, tactic is in one free
playbook linked at the top of the
description. Grab it, hit subscribe, and
we'll see you in the next
