Turn Claude Code Into a Design Genius in 3 Steps.

AI has no taste.

You already know this feeling. You open a site and within half a second you just know a model made it. Same font. Same purple gradient. Same rounded cards floating on the same soft grey. Nobody told you it was AI. You could just tell.

For a while I thought this was a model problem. Like, the models just aren’t good enough yet, give it a year. But that’s wrong. The models can get 10 times better tomorrow and it won’t fix this. All that happens is the “generic” look moves somewhere new. The problem was never the model. It’s that AI has no taste, and most of us aren’t giving it any.

I go down a YouTube rabbit hole most weeks looking for workflows that actually hold up when you put real work through them. Most don’t. Then I found one from Chase on beating this exact “AI slop” problem in Claude Code, and I ran the whole thing end to end on a dummy SaaS site, a fictional AI infra company, from the one-shot slop baseline to the finished page. It’s the first design workflow I’ve tried that didn’t quietly fall apart halfway through.

So let me walk you through it. I’ll tell you what I kept, what I changed, and where it actually matters. Because I run an MVP agency, and design is the difference between a client saying “cool demo” and a client paying you.

Here’s the full breakdown.

First, what “AI slop” actually is

Let me explain what slop really means, because it’s not what people think.

Slop is not a bug. It’s not a broken layout or a typo. The site works fine. It’s just generic. It looks like the statistical average of every website the model ever saw, which is exactly what it is. Average font. Average spacing. Average everything. Regression to the mean, rendered in HTML.

And here’s the part that matters. You can’t fix generic by prompting harder. “Make it more premium” does nothing, because the model has no idea what premium means to you. It has no reference. No taste. It’s guessing at the average of “premium” too.

The fix is simple to say and takes real work to do. You have to inject your taste into the process. Not the model’s taste. Yours. That’s the whole game, and it’s what makes the output actually yours instead of everyone else’s.

The workflow is three steps. Cultivate taste. Give Claude the right tools. Then build the right way.

The mistake almost everyone makes

Before the steps, the thing to unlearn.

Almost everyone treats Claude Code like a slot machine. You type “build me a premium SaaS landing page,” you pull the lever, you pray. If it’s bad you type “make it better” and pull again. That’s not a workflow. That’s gambling.

I did this for months. The reason it fails is that you’re asking the model to guess your taste out of thin air, one prompt at a time. You’re never going to win that way. So let’s stop pulling the lever and give it something real to work from.

Step 1: Cultivate your taste first

This is the step people want to skip, and it’s the most important one.

You cannot inject taste you haven’t built. So before you touch Claude Code, go collect a library of design you actually like. I use three places. Dribbble for polished landing pages. Pinterest for hero sections and layouts you’d never think of. And X, honestly my favorite, because the design people there are doing stuff you won’t see anywhere else.

The job is dumb and simple. Scroll, find things you like, screenshot them. Save the links too if you can. Do this until you have a folder that basically is a picture of your taste.

Now the nice part. Instead of leaving those in a random folder, have Claude Code build you a tiny web app that becomes your inspiration library. It groups the screenshots by style, labels what each design actually is, and gives you the vocabulary for it. So a screenshot isn’t just “looks cool,” it’s “voxel-rendered landscape, warm palette, heavy serif.” That vocabulary is what you feed back into the model later. And it can spit out an image prompt and a build brief straight from a reference you saved.

That’s the old way versus the new way in one move. Old way, you start from nothing and hope. New way, you start from a curated pile of stuff that’s already working, that’s already yours.

Step 2: Give Claude the tools it doesn’t ship with

Claude Code out of the box is fine. But there are a few skills and MCPs that take it from fine to actually good at design. These are the three I lean on.

Impeccable. This is the best front-end design skill in the game right now, and it’s not close. It’s open source, sitting at over 50,000 stars on GitHub, and GitHub liked it enough to make it part of their own AI tooling. It’s one skill with 23 commands. You run “bolder” and it pushes a safe design toward impact without tipping into chaos. You run “overdrive” or “clarify” and it cleans up different things. Under the hood it’s hunting slop across seven areas: typography, color, spacing, responsiveness, interaction, motion, and the actual words on the page. It even has a live mode where you click through your site component by component and fix things in real time. It’s the closest thing to designing with your hands instead of the terminal.

Taste Skill. If Impeccable isn’t your thing, this is the other one I’d trust. Version two just dropped, it’s over 68,000 stars, and it works the same way. It finds the AI tells, strengthens your layout, fixes typography and spacing, and pushes the output toward something unique instead of boilerplate. Both of these are a real step above the default Anthropic front-end skill.

Higgsfield MCP. This one gives Claude a capability it flat out does not have on its own. Making images and video. I use it for hero imagery mostly. You install it once, it connects Claude to basically every image and video model out there, and now when your page needs a hero background, Claude just generates it instead of you hunting for a stock photo that half-fits.

One more, quieter one. 21st.dev, for individual components. Buttons, pricing sections, cards. You find one you like, copy the prompt, paste it into Claude. Same idea as your taste library, just zoomed in on the small pieces.

Now the honest warning, because Chase makes this point and he’s right. Do not fall down the tool rabbit hole. It’s really tempting to think you’re one magic skill away from perfect design. You’re not. The flashy one-trick skills all give you exactly one look, which is just a fancier flavor of slop. The reason I stick with Impeccable and Taste Skill is that they’re flexible, not prescriptive. They won’t hand you a great site on their own. That still comes from your taste and your prompting.

Step 3: Build wide, then narrow

Here’s the step that changed how I do this completely.

Stop trying to one-shot. Seriously. The move is to cast a wide net first, then close in. When I start a page now, I don’t ask for one version. I ask for five, in five different styles. All at once, on one screen.

Why five? Because I genuinely don’t know what I want until I see it next to other options. Sitting in a terminal trying one thing, then another, then another, is slow and it’s blind. Show me five directions side by side and in ten seconds I know which way to go.

So I pick the one direction that feels right. Say it’s a quiet, minimal, editorial look. Then I ask for three versions of just that style, tweaking the body layout. From those three I pick one. Now I’ve gone from a blank page to a real direction in two rounds, without gambling once.

Then I bring in Higgsfield and generate the hero image. I’ll ask for four options, pick one, ask for a few color variations of that one, pick the final. Same wide-then-narrow move, just for the image.

Last piece, and it’s a good one. Instead of guessing at fonts and colors by typing “make the heading bigger” over and over, I have Claude build a tweaks bar right on the dev server, the same kind you see inside Claude’s own design tools. Font, size, accent color, motion, spacing, all as sliders and toggles. Now I’m not asking for ten rebuilds. I’m dragging sliders and watching it change live until it looks right.

That’s the whole thing. See your options, pick a direction, tweak it live. No praying.

And this isn’t a one-off. I ran the exact same workflow a second time for a fictional car brand, from the one-shot template to the finished page, and got the same jump. One prompt gave me a dark template with lime buttons and a cartoon car drawn in CSS. The workflow gave me a cinematic golden-hour hero that looks like a real automotive campaign.

When to stop tweaking

Quick one, because this trips people up.

You can tweak forever. Don’t. Once the hero lands and the body has a point of view, ship it. The sliders are there to find the look, not to keep you fiddling for three hours. My rule is simple. When I stop finding things that are clearly wrong and start hunting for things that might be slightly better, I’m done. That’s the line.

What to watch out for

A few honest caveats after running this.

The tools won’t save you if your taste library is thin. Garbage references in, average site out. Spend real time on Step 1.

The prescriptive one-shot skills are a trap. They look amazing in a demo and give you one identical look every time. Skip them.

And this isn’t magic. It’s a flexible process, not a button. None of it needs to be followed to the letter. The point is you’re steering now, instead of pulling a lever and hoping.

What this actually means

Here’s the part I keep coming back to.

For the last two years the whole conversation was “can AI even build the thing.” That question is basically dead. It can. Everyone can build now. Which means building isn’t the moat anymore.

Taste is.

When everyone can ship a working site in a prompt, the only thing separating a demo from a paid client is whether it looks like it was made by a person who cares. That’s it. At the agency, this is the exact line between “looks cool” and someone sending money. Design is no longer the last 10%. It’s the whole edge.

So if you’ve been letting Claude one-shot your pages and wondering why they feel generic, this is the fix. Build your taste. Give the model real tools. Build wide, then narrow. You’ll stop shipping slop.

We are so early. Most people still don’t even know this is a skill you can build.

TLDR

→ AI has no taste. Better models won’t fix it, they just move what “generic” means.

→ You can’t prompt your way out of it. You have to inject your own taste.

→ Step 1: build an inspiration library, then have Claude turn it into a labeled taste app.

→ Step 2: add the tools Claude doesn’t ship with. Impeccable (51,013 stars), Taste Skill (68,083 stars), Higgsfield MCP for images, 21st.dev for components.

→ Step 3: never one-shot. Five styles, pick one, three versions, pick one, tweak live with a tweaks bar.

→ Don’t chase magic skills. They give you one look, which is just prettier slop.

→ When everyone can build, taste is the moat.

Full credit to Chase for the original workflow. His video is worth the watch: https://youtu.be/7FU98O0JLHs

LFG.