Every AI-built site looks like every other AI-built site, and the model is not the reason.
"That's not the model being bad at its job. It's the model filling in a blank you left. You gave it a goal, not a design. So it reached for the average of everything it has ever seen, and the average is slop."
The fix is not a better prompt. It is three things a model cannot invent for you: a design system to copy, motion to copy, and a judge that says when the work is still ugly.
The design system, in before any code
There are libraries of real brand design systems written out as files an assistant can read — fonts, colours, spacing, button behaviour, from companies that paid a design team six figures to settle those questions. Pick the one whose feel you want and paste it in before you ask for a single page.
AI Assistants For Your Business
We build AI systems that help companies work more efficiently and save time on repetitive tasks every single day.
AI that does the work, not just the talking.
Assistants that take the repetitive part of your week off your plate — on hardware you control.
Sample sentence, not site copy. "You gave it a goal, not a design" — the left card is a goal (sell AI to businesses); the right card is this site's own actual type and spacing rules applied to the same goal.
The judge is the part we actually need
Two open-source critics turned up in the same piece. One runs more than 58 checks aimed squarely at the defaults an AI reaches for, with commands to fix them, and a browser extension that overlays the problems on the rendered page. The other describes itself as an anti-slop framework and reads the brief before generating — industry, audience, mood, how much motion suits.
Both are enormous and both are permissively licensed: roughly 77,000 and 60,000 stars, checked in mid-August. Both are plain skill files with no runtime dependency, which is why we are interested in them and not in the component library that ships alongside.
We have not run either. We verified they exist, that they are actively maintained and that the licences are what they claim. We have not verified that they are any good. Star counts are popularity, not quality, and we would rather say that than imply we tested something we read about.
Why a critic beats another sweep
You can read every stylesheet in a project and still miss a heading set black on black in a light-preferring browser, a tile carrying no styling at all, or a page that never used the shared container. One screenshot found all three in a project where two source-level sweeps had found nothing.
A source sweep only ever finds the fault you already know to look for. A judge that renders the page and reports what still looks wrong is the missing step, and a browser overlay is that step mechanised.
Two cautions we are applying
The component library in this stack ships React, Tailwind and a build chain. Our pages are hand-written with no build step, deliberately, and one of our surfaces is a single hand-written file. That is a real cost, not a preference, so that tool is out here even though it is good.
And one of the critics infers the brief — industry, audience, mood — before it generates. That is positioning, and positioning is a person's job here, decided before anyone writes. Use it as a critic, not as a strategist. A tool that quietly decides who your page is for has overridden the one decision everything downstream depends on.
What to do
Put a design reference into the brief before you ask for a page. Then render the result and let something judge it that did not build it. If you only do one, do the second.