Every AI coding agent keeps building the same website.
I got tired of seeing it, so I built a fix.
Ask Claude Code, Cursor, Codex, Gemini, whatever, to “build me a landing page.”
You already know what comes back.
Purple gradient.
Inter.
Centered hero.
Three cards.
Pill buttons.
Maybe some glassmorphism if it is feeling fancy.
That is not really the model being stupid.
It is the model doing exactly what you asked it to do with no taste constraints.
If you leave every design decision blank, the model fills those blanks with the average of the web.
And the average of the web looks like slop.
The thing I kept noticing is that there is not just one kind of AI slop.
There are three:
1. The look
Purple gradients, beige editorial templates, generic SaaS polish.
2. The structure
Centered hero, three-card grid, equal sections, nothing with a real layout opinion.
3. The copy
“Delve,” “in today’s fast-paced world,” “not just X, but Y.”
You do not fix that by asking the model to “make it better.”
That just gets you cleaner slop.
You fix it by removing the blanks before the model starts guessing.
So I built tastecheck.
It is a free MIT skill pack for AI coding agents.
The main skill does not just clean up bad UI after the fact. It interviews you before the build starts.
It asks opinionated questions like:
- What site would you actually be okay resembling?
- Are we going warm or cool?
- What is the one dominant color?
- What should this feel like to use?
- What are we absolutely not doing?
Then it turns those answers into a real design direction: type, color, tokens, layout rules, interaction states, accessibility checks, and the rest of the system the agent should follow.
The pack has 14 skills total:
- typography
- color
- layout
- responsive behavior
- component states
- forms
- dark mode
- motion
- empty states
- data visualization
- accessibility
- cognitive accessibility
- design review
- and the main taste interview
The test I care about most:
I took the same content and ran it through five different design systems.
Not five recolors.
Five different products.
Copper.
Swiss.
Maximal.
Concrete.
Clay.
Same words. Same core page. Completely different layout grammar.
That is the point.
Taste is not “make it prettier", taste is constraint.
It is deciding what the thing is allowed to become before the model averages it into mush.
It works with Claude Code, Codex, Cursor, Gemini, Kilocode, Kimi, etc. It is just Markdown. No SaaS. No runtime.
Repo:
Samples:
I built this because I do not think the answer to AI slop is “try harder in the prompt.”
I think the answer is giving the agent taste before it starts building.
Question for you:
What is the AI design tell you notice first?
The purple gradient?
The same hero/card structure?
The “delve” copy?
Or something else entirely?
12
8 comments
Simon Gonzalez De Cruz
6
Every AI coding agent keeps building the same website.
Clief Notes
skool.com/cliefnotes
What we give away free beats most paid courses. Build durable AI systems with a Marine vet and Edinburgh researcher. 40+ lessons, growing.
Leaderboard (30-day)
Powered by