The New AI prompting Method Everyone Is Talking About: Loops (And How To Use Them)

At a glance

The top AI programmers have quietly stopped prompting chatbots one message at a time. Instead they build loops: a system where you set a goal once and the AI prompts itself, discovering the work, planning it, executing it across many agents in parallel, verifying whether it hit the goal, and iterating until it does. Upgraded breaks the idea down from scratch, names the two flavors (open loops and closed loops), and then builds a real one live in Claude Code: a three agent system that grows a fictional pickleball store by building a shareable quiz, scouting content opportunities, and generating launch marketing. It closes with four everyday scenarios (a freelancer, a student, a shop owner, a YouTube creator) reworked as loops you could build yourself, plus an honest note on where the AI still needs you in the seat.

The whole point: a reader who follows this page can build the same loops. Every prompt used in the demo is reproduced below, verbatim.

The core idea: stop prompting, start looping

There is a new idea the best AI users and programmers are all talking about, called loop engineering, or just loops. Boris Cherny, who created Claude Code, and Peter Steinberger, who built a tool the video calls OpenClaw, have both said this is now how they program. They do not just prompt a chatbot, and they definitely do not write code by hand. Instead they create loops where the AI prompts itself. They call it the future. It is also, admittedly, confusing at first, which is what this video sets out to fix.

Loops are usually discussed in the context of coding, but they are not only for code. You can loop content creation, research, or teaching yourself a new skill, and the video covers all of those. The catch is that they still run through coding tools like Claude Code or Codex. You do not have to be technical to use one, though. If a terminal window scares you, that is fine.

The old way was a slow back and forth. You, the human, write a prompt and send it to the agent. It programs something, produces output, and sends it back. You read it, write the next prompt, it produces again, and you repeat forever. You are iterating, but you are the bottleneck on every single step.

The new way starts with the goal, set by the human once. From there the agents take over: they run discovery to find what needs doing, plan it into clear steps, execute (this is where they can work in parallel, spinning out 15 different agents that each do one thing), then hit a verification step where a verification agent checks whether the work accomplished the goal. If it did, ship it. If it did not, it iterates and tries again. You can even add a step after shipping where an agent looks at what shipped and asks "what should we do next?", then iterates again on its own.

The other key piece is memory. It lives outside of the conversation and keeps track of what is going on and which steps have been completed, so the agents can iterate successfully rather than forgetting where they were. A more concrete way to picture the whole thing: an orchestrator agent oversees everything, knows what is happening, and hands tasks to individual specialist agents. Each specialist runs its own discovery to ship loop, and then they all come back together to create the finished product.

Here is the same shift as a ledger.

Open loops versus closed loops

There are two types of loops.

An open loop is broad. You basically tell the agent: go out there, see what we should do, then go do it. That is powerful, because the agent can discover things you would never think of. The cost is that it burns massive tokens, because it can go in any direction it wants. Even after it finishes one direction, it will pick another and just keep going. If you work somewhere with an effectively unlimited budget, that is great. Otherwise it gets expensive fast.

A closed loop is the recommended default. You start with a bounded goal. You see the path first, so you understand what sorts of things the agents might do. There is a clear evaluation at each step, and that keeps the budget fairly normal and constrained. How much it costs still varies with how much you ask the agent to do, but it should not get out of hand.

A real world build: growing a pickleball store

To make it concrete, the video builds a loop for something an actual person would have, not a throwaway coding project. The scenario: you run an online pickleball store. You sell pickleball gear, you have some customers but not a ton, and you want to grow it. Three tasks work toward that goal.

1. Build a BuzzFeed style quiz that recommends a pickleball paddle based on your favorite Harry Potter character. Goofy and silly, yes, but those things work and are shareable.
2. Build a content research engine that goes out and finds what your potential customers are actually talking about right now, across social media, blogs, forums, and Reddit. It runs on a loop, doing this every week and getting better over time.
3. Add a growth agent that takes all of that, looks at your site, and turns it into ready to use marketing: email campaigns, social captions, and site recommendations for where to link so more people take the quiz.

That is three agents doing three jobs at the same time.

The three agent system, mapped

The goal is to grow the pickleball ecommerce site automatically without doing every task manually. An orchestrator agent owns that goal. It reads a next_steps file (its memory), delegates work, and synthesizes the outputs. The most automatic setup is to make the orchestrator itself an AI agent, which the video does later. To start, the presenter plays orchestrator by hand across three agents so you can watch each one work in its own tab. If you run an orchestrator inside the Claude Code desktop app, it will create all these agents inside a single chat, which does the same thing but hides what each agent is doing, so the separate tabs are shown for clarity.

The three specialists:

• Agent 1, the builder. Builds the Harry Potter quiz with email capture at the end.
• Agent 2, the scout. Searches Reddit forums, competitor sites, and search trends for what pickleball buyers want and are talking about. Its loop condition: keep looping until there are 3 or more fresh ideas not yet acted upon (no content created about them yet).
• Agent 3, the growth agent. Reads the quiz and the site to understand where to link people to drive them to the quiz. It checks which recommendations from the last cycle were implemented and finds new opportunities, then produces emails, site edits, social captions, and a next quiz recommendation.

The orchestrator reads all of that, creates one unified action plan, and starts acting on it. Then the loop kicks in around a single question: is the site growing? Are there fresh content ideas? Is the lead magnet pipeline full, meaning are people consistently taking the quiz? If no, keep going until it gets there. If yes, pause and wait for the next steps. All of it runs on a weekly schedule cadence that automatically triggers the orchestrator to find new content ideas, analyze the site, and draft emails, and every result feeds back to the orchestrator to close the loop.

Agent 1: the builder prompt

You need a Pro plan or higher to access Claude Code. Open a new session and paste the builder prompt:

You're the builder. Your only job is to create a Harry Potter x pickleball
personality quiz.

Deliverable: a single self contained HTML file saved to /outputs/quiz.
- 6 questions total. Each question should feel fun and slightly absurd.
- Questions should subtly map to pickleball playing styles (aggressive,
  strategic, social, defensive, etc).
- 4 possible results.
- An email capture form that appears BEFORE the result is revealed, with a
  CTA that says "See your full result."

Ask questions if needed.

That closing line, "ask questions if needed," is a deliberate habit: it lets you work with the agent to land on something you actually like. Hit start.

Agent 2: the scout prompt

While the first agent runs, press Command+N to open a new session and prompt the second agent:

You are the scout. Your job is to research real content opportunities for a
pickleball ecommerce site, completely independent of the quiz being built in
parallel.

Go out to the internet and find what pickleball players, buyers, and
enthusiasts are actually talking about, searching for, and struggling with
right now. Research sources: Reddit (relevant subreddits), search trends,
competitor sites, and YouTube (which videos are getting traction).

For each opportunity, score it on:
- Audience size
- Purchase intent (will people actually use this to buy stuff)
- Content gap (is it underserved)
- Quiz or lead magnet potential (could it become a quiz, tool, or guide)

Write a ranked list of the top 8 content opportunities to <location>. For each
one, include the topic, a one line angle, the source where you found traction,
the 4 scores, and a recommended content format. Log progress to the log files.

Those log files just track what the agents are doing, so you can look back and each agent knows where it is in the process.

Agent 3: the growth agent prompt

Press Command+N again for the third agent, the growth agent. It is meant to start acting once the other two finish, but because it runs on a loop it keeps working in parallel with them:

You're the growth agent. The quiz has been built and content research is done.
Your job is to do everything a smart marketing hire would do in the first 48
hours after a product launch. Read the quiz and research files before doing
anything. Then complete 4 tasks:

1. Site link audit. Think of a normal pickleball ecommerce site and identify
   every page or section where a link to the quiz would naturally fit. For each
   placement, write the exact location, the exact copy to add, and the reason
   it fits there.

2. Launch email. Write a launch email announcing the quiz to the existing
   customer list. Include subject line, preview text, full body copy, and CTA.
   Casual, fun, pickleball obsessed.

3. Social captions. Write 3 captions: one for Instagram, one for Reddit, one
   for a Facebook group. Each native to its platform. The Reddit one should NOT
   read like marketing. The Instagram one needs a hook in the first line. Save
   all 3 to /outputs/social_captions.

4. Lead magnet recommendation. Based on the top opportunities and content
   ideas, recommend the single best next lead magnet to build. Include title,
   format, a 3 sentence description of what it does, and why it will outperform
   or complement the Harry Potter quiz. Save to /outputs/next_quiz_recommendation.

Then loop: after completing all 4 tasks, evaluate. Did the site have obvious
placement opportunities that were missed in a previous cycle? Are any of the
social captions similar to ones written in a prior run? Flag any repetition or
diminishing returns in /outputs/growth_agent_notes. This keeps future cycles
fresh. When all tasks are complete, log "growth agent complete."

Note where the site link audit sidesteps a missing asset: there is no real website in the demo, so the prompt tells the agent to imagine a normal pickleball site and reason about placements. That is a small trick worth copying whenever the real thing does not exist yet.

The results

The builder one shot the quiz and it was live. Titled "The Sorting Paddle," it walks through absurd questions ("You show up to open play and the vibe is chaotic, balls flying everywhere, someone's blasting EDM, what do you do?" with an answer like "I observe for a few minutes to identify who the real players are"), gates the result behind an email capture ("Enter your email to reveal your Hogwarts house playing style breakdown and your perfect paddle match"), and returns a house plus a paddle recommendation. It is not perfect and has a bit of that AI look, but it works and would be a nice add for a pickleball site.

The scout returned a ranked list of opportunities: "is my paddle legal?", an intermediate upgrade trigger guide, an injury prevention gear guide, shoes and when the upgrade pays off, women's gear content, and a beginner starter kit quiz. Solid at a glance.

The growth agent produced a full post launch marketing strategy. Task 1 alone was 12 placement recommendations with copy for each (homepage, sticky navbar, category pages), for example placing the quiz directly beneath the main homepage headline with copy like "Which Hogwarts house matches your pickleball personality?" (it lacked the full quiz context so it got a detail wrong, honestly flagged). It also drafted the launch email with copy notes, platform native social captions (the Reddit caption opens "I've been playing for about 3 years now and I was bored the other night thinking about why certain player personalities are so consistent," which reads like a person, not an ad), and a next quiz recommendation, "Is your paddle past its prime?", complete with why it would outperform and supporting data points.

Automating it: the orchestrator agent

Watching three agents by hand is instructive, but the payoff is not being the orchestrator yourself. Replace yourself with an orchestrator agent:

You're an orchestrator agent managing a fleet of 3 sub-agents for a pickleball
ecommerce business. Your job is to delegate work, monitor outputs, and
synthesize results into a unified action plan.

Before doing anything, check if /outputs/next_steps exists. If it does, read it
FIRST. It is your memory from the previous cycle. Use it to understand what is
already completed and what still needs work before delegating anything.

Goal: grow the pickleball ecommerce site by launching a new lead magnet,
researching content opportunities, and executing a 48 hour post launch
marketing push.

Spawn the 3 agents (builder, scout, growth), each with the exact prompt used
above. Then synthesize their outputs and write /outputs/next_steps: a summary,
the top 3 actions to take this week, and what the next loop cycle should focus
on.

Loop condition, evaluate all 3:
- Are there at least 3 unacted content ideas in the document?
- Is the site fully linked to the quiz?
- Is the next lead magnet defined?
If any condition is not met, keep going.

The orchestrator spawns each agent with the same prompts shown earlier, then does the part that makes it a loop: it synthesizes everything and writes the next_steps file (a summary, the top 3 actions for the week, and what the next cycle should focus on). Writing that file is where the looping begins, because the next run reads it first. If the loop conditions are not met, it keeps going. From here you can spin out even more agents: a copywriter agent that turns each content idea into a blog post and cycles back in, and once all the ideas are written, new ones get spun out. That is how the system starts building on itself.

Four scenarios to loop yourself

The video ends with a quiz for you. Four real world scenarios, each one a candidate for a loop. The exercise is to think through how you would build a team of agents to loop it before reading the suggested answer.

1. The freelancer. You are a freelance designer with 6 active clients. Every Friday you spend 2 hours writing status updates, digging through folders, summarizing progress, and personalizing each message. Same task, every week, and you hate it.

Loop it: a loop runs every Friday at 4 PM automatically. It reads your project folders to see what changed, loads a skill file that knows each client's name, project goals, and preferred tone (the memory of who your clients are and how you write), drafts a personalized status update, drops the finished draft in a review folder so you can look it over (keeping a human in the loop for quality assurance), and logs what was sent so you never repeat the work. Loop condition: did every active client get an update this week? If yes, done. If not, keep going.

2. The student. You are studying for exams and trying to stay current in a fast moving field like AI. New papers drop daily, new tools weekly. You feel behind and cannot read everything, but you do not want to spend 3 hours every Sunday on research.

Loop it: it loops every Sunday while you sleep, searches for the 5 biggest developments of the past week in your topic, scores each by relevance to what you are actually studying, and filters out anything below your relevance threshold. It writes a plain English briefing (what happened, why it matters) and checks the past 3 weeks of briefings so it does not repeat itself. Loop condition: are there 2 or more genuinely new developments this week? If not, dig deeper.

3. The online shop owner. You sell products online with lots of SKUs. You know you should update descriptions, fix what is not converting, and push your winners harder, but it takes too long so you never do it.

Loop it: run it once a month, on the first. It reads your sales data (what sold, what did not), identifies the 3 products with high traffic and low conversion, and rewrites those descriptions with better hooks and clearer CTAs. It writes promotional copy for your top 3 performers to push them harder, and logs every change made and the reason behind it.

4. The YouTube creator. You post once a week and have a pile of video ideas, but never know which to make next, so you pick on vibes. Sometimes it works, sometimes it does not.

Loop it: a loop runs every Monday morning before you start work. It reads the full ideas list, pulls the last 90 days of video performance, identifies which topics overperformed and which flopped, checks what is trending, scores every idea, and outputs a ranked top 5, while flagging ideas competitors have already covered.

Where AI falls short (and staying in the loop)

As much as the presenter loves AI, he will not pretend it is always perfect. For each scenario you have to understand what the AI is good at and what it is not. The YouTube example is his own case in point: as someone who has spent a lot of time making videos, he knows the AI is not always right about which topics are genuinely good. So you think for yourself and think critically. Even so, having the loop surface a batch of ideas and a second opinion is genuinely helpful. Build it, then add your own judgment on top. The takeaway is not blind automation, it is a human owning the goal and the final call while the loop does the grinding.

If you have a Claude subscription or a Codex subscription, you can start testing this right now. The presenter frames it as the future of how people will work with AI.

Key takeaways

• Stop prompting one message at a time. Set a goal once and let the AI loop. The loop runs discovery, planning, parallel execution, and verification, iterating until the goal is met.
• Memory lives outside the chat. A file like next_steps records what is done so cycles compound instead of repeating. The orchestrator reads it first, every run.
• Use a closed loop by default. Bound the goal, see the path, evaluate each step, and the cost stays sane. Save open loops (broad, self directing, token hungry) for when budget is no object.
• Structure work as an orchestrator plus specialist agents. One agent owns the goal and delegates; each specialist runs its own loop with a clear loop condition (for example, "keep going until 3 or more fresh ideas exist").
• Write explicit loop conditions and log everything. The condition is how the system knows when to stop or continue; the logs are how the next cycle avoids repeating itself.
• Give each agent a single job and a "read these files first" instruction so parallel agents stay independent and coordinated through shared memory.
• Add "ask questions if needed" to collaborate with an agent toward output you actually like.
• It is not only for code. Loops apply to content, research, marketing, and personal workflows, though they still run through tools like Claude Code or Codex.
• Keep a human in the loop. Route drafts through a review folder and keep the final judgment yours, because the AI is not always right about what is actually good.

Chapters

0:00 Why the top AI programmers stopped prompting 0:44 What a loop actually is (the old way versus the new way) 1:28 The 5 stages of a loop, explained visually 2:45 Open loops versus closed loops (and which to use) 3:35 Real world example: growing a pickleball store 5:00 The 3 agent system, mapped out visually 7:16 Building it live in the Claude Code desktop app 7:28 Agent 1: the quiz builder prompt 8:07 Agent 2: the content scout prompt 9:21 Agent 3: the growth marketing agent prompt 11:18 Checking the results 13:38 How to automate it all with an orchestrator agent 15:35 QUIZ: how would you loop these 4 scenarios? 15:47 Scenario 1: the freelancer 16:39 Scenario 2: the student 17:22 Scenario 3: the online shop owner 17:53 Scenario 4: the YouTube creator 18:22 Where AI falls short (and why you stay in the loop)

Notable quotes

"They don't just prompt a chatbot, and they definitely don't write code by hand, but instead they create these loops where the AI prompts itself." (0:20)

"You start with the goal. This is set by the human once." (1:31)

"It burns massive tokens because it can go in any direction that it wants." (2:52)

"So say if you work at Meta and you have unlimited budget, then this is great. But otherwise, I would recommend the closed loop." (3:11)

"This was one-shotted. I think this would be nice for a pickleball website. Obviously it's not perfect. It's got the kind of AI look, but you guys get the point." (11:38)

"Before doing anything, check if /outputs/next_steps exists. If it does, read it first. It is your memory from the previous cycle." (14:05)

"As much as I love AI, I don't want to pretend like it is always perfect." (18:22)

"It's extremely helpful and it's the future of how people are going to work with AI." (18:52)

Resources mentioned

• Claude Code by Anthropic, the coding tool the loops are built in (requires a Pro plan or higher).
• Codex, OpenAI's coding agent, named as the other tool loops run through.
• Claude, the subscription that unlocks Claude Code.
• Boris Cherny, creator of Claude Code, cited as an early loop advocate.
• Peter Steinberger, creator of the tool the video calls OpenClaw, also cited as a loop advocate.
• Reddit, a primary research source for the scout agent.
• YouTube, a research source for content traction and the creator scenario.
• BuzzFeed, the reference for the shareable personality quiz format.
• Meta, used as the example of an unlimited budget where open loops make sense.
• Upgraded, the channel that produced this walkthrough.