Turn Claude Into A Design GENIUS In 3 Simple Steps

At a glance

Chase Hannegan has a specific enemy in mind and he names it in the first sentence: AI slop. Not the broken kind, the generic kind. The website that works fine and still announces, from across the room, that a model made it. His argument is that this is not a technical problem and no amount of model improvement will fix it, because a ten times better model just moves where the average sits. The fix has to come from you. So he lays out three steps: cultivate your own taste and store it somewhere Claude can read, arm Claude Code with the handful of external skills and MCPs it does not ship with, and stop trying to one shot the build. The third step is where the video earns its length. He builds a landing page for a fake analytics startup called Kestrel live, in four widening passes: five aesthetics, then four body layouts, then four Higgsfield generated hero images, then five color grades of the winner, and finally a live tweaks bar so he can move fonts and accents without asking Claude anything. Every prompt is read out loud. Every rejected option is named.

Chapters

• 0:00 Intro
• 0:59 Step 1
• 3:55 Step 2
• 9:58 Step 3
• 22:05 Outro

The two faces of AI slop

The video opens on a before and after that is really a before and before.

The first image is the slop everyone already learned to spot: a generic SaaS template, Inter everywhere, blue purple gradients (0:10). That was the tell for a couple of years. Purple to blue, rounded cards, a floating 3D blob, done.

The second image is the current version, and this is the interesting one. "This is definitely better, but it still screams AI" (0:20). Nothing is wrong with it. The type is fine, the spacing is fine, the layout works. What gives it away is that it has a specific color palette and a specific font and a general style, and you can tell from across the room that it came out of one or two prompts.

That distinction sets up the whole video. The problem is not that models produce bad work. It is that they produce the average of good work, and the average has a look.

He promises three things: how to cultivate and inject your own taste into the web design process, which skills and MCPs to start using, and the exact build sequence, meaning what to prompt, how to prototype, and what it actually means to iterate until you get something you like.

Step 1: AI has no taste, so bring your own

"AI has no taste" (1:00) is a phrase he expects you have heard, and he thinks it is worth taking literally, because it defines what step one is for. How do we cultivate taste, how do we curate it, how do we bring it into the web design process.

Then the load bearing claim (1:10):

The problem with this website and this website isn't a technical one. It's simply generic. And no matter how good the models get, they could get 10 times better tomorrow. It'll just change what we consider to be generic.

That is worth sitting with. It says model progress cannot solve this. Better models raise the floor, and the floor is exactly what "generic" means. Whatever the best default output is, that becomes the new average, and the new average becomes the new tell. The only escape is to hand the model something that is not the average, and the only thing you have that qualifies is your own preference.

The payoff of doing that is two things at once. Something unique, because it is specific to you and therefore subjective. And something good, in the quote unquote sense he is careful to flag, because you sourced it from work that is already working.

Where to look

The instruction is unglamorous: go collect. He names three places and shows each one.

Dribbble (2:05), which he pauses to spell, "that's Dribbble with three b's." He types web design, sorts by popular, and gets a wall of landing pages.

Pinterest (2:13). Same query, web design, and what comes back is a set of hero pages that look nothing like what Claude produces when you say "build me a landing page."

X, which he still calls Twitter (2:22), and which is his favorite of the three. The reason is people rather than boards: "there are a ton of awesome creators in this space that are doing really cool design work and UI stuff that you just aren't going to see anywhere else."

Your first job is the simplest one in the video. Go to these sites, look for things you like, and screenshot them. Save the links too when the live site exists, since a URL carries motion and interaction that a screenshot throws away.

Turn the screenshots into something Claude can read

You can stop at a screenshots folder. He did something better, and this is the single most reusable idea in step one: he had Claude Code build him a small web app that is the inspiration library (2:50).

It does more than store. It groups the screenshots by design type, and it can explain what a given design actually is: what it means, and what the vocabulary associated with it is. That last part matters more than it sounds. Design has names for things, and if you cannot name the thing you like, you cannot ask for it.

Click any screenshot in the app and you get:

• Keywords for that design. The example he opens is tagged as a rendered landscape look.
• Copy image prompt. For when you want to generate an image in that style, typically the hero background.
• Copy brief. For when you want to build an entire website in that direction. This is the button that step three consumes.

So the library is not a mood board. It is a prompt source. Every entry has already been translated from "I like this" into text an agent can act on, which is exactly the translation most people never do.

He closes step one with an honest note about why the library exists at all (3:40):

So you don't just have to sit there and come up with something from scratch, come up with something that's been sitting in your brain. Because when we're just getting started, we're really not good enough for that yet. We need to find stuff that's actually working that we like and use that as our base.

Step 2: the tools Claude Code does not ship with

Step two is everything external you bolt on so that step three starts from a better place. He covers two skills, one MCP, and one website that is neither.

Impeccable

Impeccable is, in his words, "the best front-end design skill in the game right now" (4:04). The facts he gives:

• Open source, on GitHub, with almost 50,000 stars at the time of recording.
• GitHub has since made it an official part of GitHub's AI tool, meaning it ships built in rather than as a third party add on. (See the honest footnote at the end for exactly which GitHub product that is.)
• It is one skill containing 23 commands, each of which improves a design or a component in a different direction. Critique the issues. Add polish. Make it bolder. Make it quieter. And nineteen more.

He is upfront that a list of 23 command names is useless, so he opens impeccable.style where all 23 are in the left rail with a live before and after for each (4:42). Three he walks through:

• bolder, described on the site as pushing safe designs toward impact without sliding into chaos. The left panel is a plain Claude Code output. The right panel is the same thing after impeccable bolder.
• overdrive, same treatment, standard Claude Code on one side and Impeccable on the other.
• clarify, which he demonstrates on warning copy: the standard warnings you get, next to something condensed and actually easy to understand.

The frame he lands on is that Impeccable is fundamentally a slop detector with opinions. It works across seven areas (5:20): typography, color, spatial design, responsiveness, interaction, motion, and UX writing. The site carries a whole section devoted to slop, broken into 46 different patterns, and the Impeccable CLI will walk your codebase looking for them and surface what it finds on a dev server (5:37).

Then the feature that closes the gap with visual tools: live mode (5:45). Bring your site up, click through it component by component, and make adjustments right there. His comparison is direct: "it's almost like Claude Design in one sense, but it definitely gives you a visual aspect that you can't get if you're using exclusively the terminal."

Installing is deliberately boring. There is a CLI, three install options on the site, and if any of that confuses you, copy the page URL, paste it into Claude Code, and let the agent install it for you (6:03). That last trick generalizes to almost every tool in this video.

The Taste Skill

If Impeccable is not for you, his alternative is the Taste Skill (6:16). It had just released version two, flagged experimental, and sat at just under 66,000 stars on GitHub when he recorded.

It works the same way in principle: hunt the AI slop tells, then push in the other direction. Stronger layout. Better typography, motion, and spacing. The stated goal is something unique rather than a boilerplate looking UI.

His verdict on the pair (6:43): "I think these two, the Taste Skill and Impeccable, are the best in the game right now. Definitely a step above the standard Anthropic frontend design, and definitely a step above some of the more popular repos like UI UX Pro Max."

That is a pointed ranking. UI UX Pro Max is the most starred of the four by a wide margin and he still puts it below the other two, which is a preview of the argument he closes step two with.

The Higgsfield MCP

MCPs get one entry and it is the one he uses most: Higgsfield (6:53). The reason is a plain capability gap. Claude Code cannot make images or video. Higgsfield's MCP hands it both, and it does so as an aggregator rather than a single model, which he describes as access to "virtually every single AI image and video generator out there."

His current picks inside it (7:10):

• Images: GPT Image 2.
• Video: Seedance. He flags this one as unstable in the sense that matters, "that's usually bouncing around every few weeks," but right now Seedance tends to be the best.

Installing follows the same pattern as everything else. Go to higgsfield.ai, open MCP and CLI, copy the snippet, paste it into Claude Code, and it walks you through authentication (7:22). There is a CLI as well as the MCP, and the site also offers skills you can add.

Once it is wired, Claude Code just reaches for it. If you are building a site that wants hero imagery in the background, the agent calls the MCP and generates it. Same for any other custom asset. Step three demonstrates exactly this.

21st.dev

The last item in step two is neither a skill nor an MCP: 21st.dev (7:52). He notes it arguably belongs in step one, since it is really a taste source, just at a lower altitude. Step one is about big picture aesthetics. This is about components.

The workflow he shows: click buttons in the left rail, get a wall of buttons, click one you like, and there is a copy prompt button underneath it (8:20). Paste that prompt into Claude Code and you get that button. Not a screenshot of it, the actual thing.

It is not just buttons. Cards, pricing sections, borders, backgrounds, calls to action. And his favorite example of why this is worth browsing at all is the least glamorous element on the internet (9:03):

How often are you looking at like pagination and the different options we have for that?

Which is the same argument as step one, scaled down. You cannot ask for a component you have never considered. Browsing a registry is how you find out that a decision you have been defaulting on was a decision.

The rabbit hole warning

Step two ends with a caution that is really a rule for evaluating any design tool (9:05):

Beware going down the rabbit hole in terms of tools and skills, because it's very easy to fall into the trap of, well, I'm just one skill away from all my design problems being solved.

You will find skills that build genuinely cool websites. The catch is that they tend to be narrow in scope and extremely prescriptive, and a prescriptive skill only ever gives you one kind of output. Which is a fine deal if that output is what you want and a dead end otherwise.

The reason he picks Impeccable, the Taste Skill, and Higgsfield is that all three are flexible, and you can push them in many directions. And then, to his credit, he immediately names the cost of that choice (9:36):

However, because they are not prescriptive, they aren't always going to give you a quote unquote great output, because that great output is going to depend on your prompting and again how we inject taste.

That is the honest trade. A prescriptive skill guarantees a decent floor and caps the ceiling. A flexible skill removes the cap and removes the floor with it. Step three is the procedure for surviving without the floor.

Step 3: the build phase

The thesis arrives immediately (10:00): "what I want you to get away from is the idea that we need to one shot these things." What replaces it is cast a very wide net off the bat.

He shows the demonstration he ran before recording. A website for a fake AI company called Argus, built in five different versions in five different styles, all displayed at once. And there is a second axis in that grid: the left hand column used Impeccable, the right hand column used the Taste Skill (10:17). One screen, ten data points, two design skills compared head to head under identical prompts.

The cadence he recommends (10:31):

1. Five styles. "I like to do five." Once you see all five you have a decent idea of the direction you want.
2. Three versions of the winner. Take the one you picked, ask for three iterations of that style.
3. Pick one and tinker.

The reason for doing it this way is spatial, not technical (10:48):

The idea is I want to see all my options on one screen at one time. It's kind of tough, I think, to know what direction you even want to go on when you're just sitting inside a terminal and you just try one thing and then you try another. Like, why don't we just try a bunch of stuff at once and compare and contrast?

Assets come last. Only once you have fidelity on the style do you bring in the Higgsfield MCP and start generating hero images and the rest (11:12). Generating a beautiful hero for a layout you are about to throw away is wasted work.

The four things every prompt carries

Prompting gets a short, specific framework, and he opens it with a shot at the alternative (11:31):

It's not some 10,000 page design.md prompt you need to copy, because remember, when people give you that, it's going to be kind of the same thing every time. It's very narrow.

Which is the same complaint as the prescriptive skills warning, applied to prompt libraries. A giant shared design document is just prescriptiveness in another format, and it converges everyone who uses it onto one look.

His four slots:

Two notes he adds to slot two. First, the point of a reference is to match the feel, not the content (12:02). "We aren't necessarily trying to match it contentwise or design-wise, because we're not trying to like copy them. Again, we're going for a feeling, an aesthetic." Second, references do not have to be images. Drop actual website URLs, because Claude Code can go and look at them.

Slot four is where the AI slop vocabulary earns its keep. "I never want purple gradients. I never want Inter font" (12:38) is only a usable instruction because he has spent enough time in step one to know what the tells are called.

Prompt one: five aesthetics at once

If you built the library, you can offload the aesthetic selection to Claude (13:26). His is already broken out into families, print tech paper, dither mono, the vast quiet cinematic thing, so he can say: take a look at my library and pick out five different aesthetic families and create websites like that. If you have no library, "I want five different variations in five different styles" works too, it just leaves the taste up to the model.

The prompt he actually ran, as he reads it out (13:47):

Build a landing page for Kestrel, an AI analytics platform for small startups.

Intent: get the visitor to book a demo.

Guard rails: every hero should carry a monumental image.
Never use AI slop patterns: no purple gradients, no 3D SaaS blobs, no Inter font.

Create five versions of the page. For each version I'll give you:
  - the aesthetic (the design family it should follow)
  - a reference image from my inspiration library
  - what the future hero image will probably look like

The five came back and he walks the grid (14:28). Print tech paper, with what he calls data as texture. Vast quiet, holding a placeholder for a mountain range that does not exist yet. Dither mono. Classical remix. Four of the five get named clearly on screen.

And a detail worth flagging, because it is easy to miss (14:37):

Remember, all these graphics were all just generated by Claude Code. I haven't used anything like the Higgsfield MCP to actually create the graphics.

So pass one runs entirely on Claude Code's own drawing. The image model has not been touched yet. That is the discipline he described earlier, actually executed.

His pick: dither mono looks cool, but vast quiet wins (15:10). "It's definitely different. Very minimalist, but I kind of like what it's doing. And I think this might look really cool if we can nail the actual image behind it."

Prompt two: three body layouts

Same move, narrower (15:33):

Let's go with the vast quiet version. Generate three versions of that aesthetic
for me, namely changing the body formats.

He notes you can be as specific as you like about what should differ between them. He chose body format because the hero direction was already settled and the body was the open question.

Four panels come back, the original plus three new ones (15:57):

• Original. Very minimal, very vertical.
• A second calm one, same approach, with more going on in the body.
• Ledger. Things shift left and a scrolling index follows you down the page.
• Frames. Edges on the sides that frame each section.

He picks the ledger (16:30). "I kind of like how when you scroll down this follows you on the left hand side. Maybe we add something to the right a little bit. Not totally sure, but I like this. This kind of is like a neat, minimal approach, yet looks a little bit different." It gets the label V3B, the quiet ledger version.

Prompt three: nail the hero image

Now, and only now, the image model gets called. The sequencing is deliberate (16:50): "I want to nail the hero image first. I think if I nail this background, then I'll begin playing around with fonts and where things are kind of sitting."

The prompt (17:03):

Let's go with V3B, the ledger version.

Now nail that hero image. You have the examples of hero images that are part of
the quiet aesthetic in our inspiration library, I think there are three or four
of them. Give me four different images that fit that, and that would also fit
our hero composition specifically.

Use the Higgsfield MCP. Make them high quality, make them 2K, create four
different designs, and pull them up once you're done.

Note what that prompt is doing. It is not describing a picture. It is pointing at the library entries that already encode the aesthetic, adding a compositional constraint from the chosen layout, and naming the tool, the count, and the resolution. Four separate constraints, none of them adjectives like "premium."

Claude Code called the Higgsfield MCP, generated the images, dropped them into the hero section, and displayed the variants (17:42): aerial, crag, watercolor, cloud sea.

His read: not a fan of the second, "I feel like it overlaps some of the text," and he likes one and four. He wants one thing changed. It is very black and white and he wonders whether a splash of color somewhere would help.

Prompt four: color grade the winner

Can we go with option four that you created, and can you create multiple
versions of that and show me just like you did before, but could we add maybe a
little color to it? Right now it's very black and white, I'm not sure if just a
splash of color somewhere would look good.

Five come back (18:35): the original, dawn, golden hour, alpenglow, and a duotone he dismisses on sight as "way too much."

Alpenglow wins.

Prompt five: assemble it, with weight

With the hero settled, he goes back to the page, and this prompt is where the taste vocabulary shows up as motion rather than color (18:53):

Let's go with the alpenglow version. Bring up our web page with that as the hero.

When you do that, take a look at the transitions between the hero and the body.
It shouldn't just be a super sudden change, do what you need to make that look
premium.

And in terms of page loading, we want everything to feel kind of heavy, give
some weight to it.

The result (19:12): elements load in one at a time, which he likes. Scrolling down, the hero does not hard cut, it slowly fades out. The scrolling index is still his favorite part, "I really like this index, honestly, I think it looks really cool." Book a demo sits at the bottom, closing the loop on the intent slot from the very first prompt. Overall very minimal, and he thinks it fits the vibe.

The tweaks bar, or how to stop guessing

Here is where he wants a bit more punch, and here is where most people would type "make it look more premium." He calls that out as the failure mode (19:52) and replaces it with something structural: have Claude build you a tweaks bar, the same live control panel you get inside Claude Design.

The prompt (20:02):

Can we sort of mimic what happens inside of Claude Design and add a tweaks bar
that pops up on this dev server so I can change a number of things? Whether
that's font size, font type, accent colors, basically any place where you think
there's a decision to be made in terms of the overall aesthetic and design,
especially when looking at the body, because I do like the hero.

Go pretty aggressive with what you offer me.

While that builds, he circles back to references with a point he had not made yet (20:29). Everything about inspiration so far has been about the hero, the first thing you see. The same applies to the body. Give it screenshots of bodies you like. Better, give it live URLs, "because Claude Code can see it, and I want it applied to my page." And build a second library for body layouts exactly as you built the first one for heroes.

The tweaks bar arrives as a button in the bottom right (21:08). Open it and you can change the heading font, kill the italics, change the size. The rationale is the same one that drove the five-at-once approach, just at a finer grain (21:20):

This is all built on the idea that I need to see the differences. I don't know what font's going to look good. I don't want to sit here and ask Claude Code to then create 10 different versions of the web page with different fonts. Just give me some sort of thing like this. Being able to iterate visually very quickly is what allows us to eventually get something we like.

He did ask for aggressive, and he got it. The panel exposes hero imagery, assets, motion, weight, and reveal distance on top of the typography and color controls (21:44). Reveal distance is a nice tell that the model took "go aggressive" seriously, since that is the kind of parameter you would normally never think to expose.

From there it is just tweaking, repeatedly, showing it more references of things you like, until you land somewhere you are happy with.

The closing argument

The outro (22:05) refuses to turn any of this into a recipe. The whole video has been about giving you a flexible workflow. "None of this is meant to be prescriptive. None of this needs to be followed to a tee." What it is instead is a set of general rules for answering three questions: what do I actually like, how do I get that taste to the model, and what tools does Claude Code not come with that would help.

And the last line is the one that names what all of it is for (22:30):

How do I tweak, how do I iterate, so I'm not just at the mercy of Claude Code one shots and just playing the lottery game of putting in a random prompt saying "please make it look more premium" and praying that the output is good?

That is the actual deliverable. Not a design skill, not an MCP, not a prompt. A way to stop buying lottery tickets.

He closes by pointing at Chase AI Plus for more Claude Code guides, particularly his Claude Code masterclass, linked in the pinned comments.

Best quotes

"I'm going to teach you how to defeat the monster known as AI slop in three simple steps." (0:00)

"A generic SaaS template with Inter font everywhere and blue purple gradients." (0:10)

"This is definitely better, but it still screams AI. It has a specific color palette, a specific font, just a general style that right away you can tell this was generated in just a prompt or two." (0:20)

"The problem with this website and this website isn't a technical one. It's simply generic. And no matter how good the models get, they could get 10 times better tomorrow. It'll just change what we consider to be generic." (1:10)

"Instead of relying on nothing and instead just getting an average regression to the mean output." (2:00)

"That's Dribbble with three b's." (2:10)

"Twitter, which I think is my favorite of the bunch, because there are a ton of awesome creators in this space that are doing really cool design work and UI stuff that you just aren't going to see anywhere else." (2:22)

"So you don't just have to sit there and come up with something from scratch. Because when we're just getting started, we're really not good enough for that yet. We need to find stuff that's actually working that we like and use that as our base." (3:40)

"It pushes safe designs towards impact without sliding into chaos." (4:50, quoting Impeccable's own description of bolder)

"It's almost like Claude Design in one sense, but it definitely gives you a visual aspect that you can't get if you're using exclusively the terminal." (5:52)

"If you get confused, just go ahead, copy the URL and drop that into Claude Code and it will install it for you." (6:08)

"Definitely a step above the standard Anthropic frontend design, and definitely a step above some of the more popular repos like UI UX Pro Max." (6:47)

"How often are you looking at pagination and the different options we have for that?" (9:03)

"It's very easy to fall into the trap of, well, I'm just one skill away from all my design problems being solved." (9:07)

"Because they are not prescriptive, they aren't always going to give you a quote unquote great output, because that great output is going to depend on your prompting." (9:36)

"What I want you to get away from is the idea that we need to one shot these things." (10:00)

"The idea is I want to see all my options on one screen at one time. Why don't we just try a bunch of stuff at once and compare and contrast?" (10:48)

"It's not some 10,000 page design.md prompt you need to copy, because when people give you that, it's going to be kind of the same thing every time. It's very narrow." (11:31)

"We're trying to match the feel. We aren't necessarily trying to match it contentwise or design-wise, because we're not trying to copy them. We're going for a feeling, an aesthetic." (12:02)

"I never want purple gradients. I never want Inter font." (12:38)

"Remember, all these graphics were all just generated by Claude Code. I haven't used anything like the Higgsfield MCP to actually create the graphics." (14:37)

"It's definitely different. Very minimalist, but I kind of like what it's doing. And I think this might look really cool if we can nail the actual image behind it." (15:10)

"The duotone, which is way too much." (18:42)

"We want everything to feel kind of heavy, and give some weight to it." (19:08)

"This is all built on the idea that I need to see the differences. I don't know what font's going to look good." (21:20)

"Being able to iterate visually very quickly is what allows us to eventually get something we like." (21:34)

"So I'm not just at the mercy of Claude Code one shots and just playing the lottery game of putting in a random prompt saying 'please make it look more premium' and praying that the output is good." (22:30)

The method, extracted

For anyone who wants the thing without the walkthrough, here is the whole video compressed into a sequence you can run today.

• OnceSpend an afternoon on Dribbble, Pinterest and X. Screenshot anything you like. Save the live URLs too. Do not curate yet, just collect.
• OnceHave Claude Code build you an inspiration library web app. It must group by design type, name the aesthetic, list keywords, and expose a copy image prompt and a copy brief button per entry. Build a second one for body layouts, not just heroes.
• OnceInstall Impeccable or the Taste Skill, or both if you want to A/B them. Install the Higgsfield MCP for image and video. Bookmark 21st.dev for components. Stop there. The next skill is not the answer.
• Pass 1Prompt with all four slots: aesthetic, reference image, intent, guard rails. Ask for five versions in five styles, all rendered at once. No image model yet. Pick a direction.
• Pass 2Ask for three variants of the winner, changing one specific thing, usually the body format. Pick one. Give it a version label so you can refer back to it.
• Pass 3Now call the image model. Point at the library entries that carry the aesthetic, add the compositional constraint from your chosen layout, and ask for four heroes at 2K. Pick one.
• Pass 4Grade the winner. Multiple color variants of one image, shown side by side. Pick one.
• Pass 5Assemble, and specify the transitions and the loading weight explicitly. "Heavy" and "not a sudden change" are instructions. "Premium" on its own is not.
• ThenAsk for a tweaks bar on the dev server, and tell it to go aggressive. Font, size, accent, motion, reveal distance. Iterate visually instead of re-prompting.

An honest footnote

Three things are worth checking against the record, none of which changes the method.

The GitHub claim is about GitHub, not Anthropic. He says Impeccable was "turned into like an official part of GitHub's AI tool" (4:14), which is easy to hear as an Anthropic endorsement given everything else in the video is Claude shaped. It is not. Impeccable ships as a built in skill in the GitHub Copilot app, which GitHub announced on X. Anthropic's connection runs the other way: Impeccable's own README credits Anthropic's frontend-design skill as its starting point, which makes the "step above the standard Anthropic frontend design" line a statement about a fork outgrowing its origin rather than a competitive dig.

The counts move. Star counts and rule counts in a video like this are timestamps, not facts. Impeccable's slop catalog was 46 patterns when he recorded and has since grown past that. The seven dimensions are the stable part and the part actually worth internalizing, since they are a checklist you can run in your head without any tool installed. Same for the star numbers on all four skills. They were true on the day and they will drift.

"Tweaks" is a description, not a product name. The behavior he is copying from Claude Design is real and well documented, live sliders and color pickers embedded into a running prototype for theme color, accent, type size, spacing and timing. Anthropic's own materials call these adjustment knobs and layout controls rather than a capitalized Tweaks panel. It does not matter for the prompt, since you are asking Claude to build the behavior rather than to import a feature, but it is worth knowing the word is community shorthand.

One more thing that is not a correction so much as an observation. The most transferable idea in the video is not any of the tools, all of which will be replaced. It is the library with copy prompt buttons. Almost everyone collects design inspiration and almost nobody translates it into text an agent can act on, so the collection sits in a folder doing nothing. Doing that translation once, at collection time rather than at prompt time, is what makes the four slot prompt cheap enough to run five times in parallel. Everything downstream of step one depends on step one having been done properly.

Resources

The creator

• Chase AI on YouTube, the channel.
• chaseai.io, his site and consultancy. He also published a written Impeccable walkthrough that goes deeper on the skill than the video does.
• Chase AI Plus on Skool, the paid community with the Claude Code masterclass he mentions in the outro.
• The free Chase AI community.

Step 1: taste sources

• Dribbble. Search web design, sort by popular.
• Pinterest. Search web design, look at hero pages.
• X, his favorite, for creators posting UI and design work that never reaches the aggregators.

Step 2: skills, MCPs and registries

• Impeccable, his top pick. GitHub repo. One skill, 23 commands, a CLI, and a live browser mode. Its slop catalog is browsable on the site.
• Taste Skill, the co-recommendation. Version two was newly released and marked experimental. Site.
• Anthropic's frontend-design skill, the stock baseline both picks are measured against, part of the Anthropic skills repo.
• UI UX Pro Max, the popular repo he explicitly ranks below the other two. Site.
• Higgsfield, the image and video MCP. The MCP page and the CLI page are where you copy the install snippet.
• 21st.dev, the component registry with copy prompt buttons on every button, card, background and pricing block.

Models named

• GPT Image 2 from OpenAI, his default for images inside Higgsfield.
• Seedance from ByteDance Seed, his current pick for video, with the caveat that the best video model changes every few weeks.

Anthropic surfaces

• Claude Code, the harness everything runs inside, from Anthropic.
• Claude Design, the source of the tweaks bar idea. Announcement and getting started guide.

Named in passing

• Inter, the font that became a slop tell through no fault of its own, and one of his standing never rules.