Anti-slop skills now have their own slop.
The recipes written to stop AI from designing generic pages are being applied word for word. What was meant to be the way out of slop is now its next tell.
Once the purple gradient became a joke, the obvious fix showed up: skills for coding agents that tell the AI what not to do and what to do instead. The best known is taste-skill, MIT-licensed. Its first version and its variants, now kept in taste-skill-classics, didn't just ban things. They handed out exact recipes.
The trouble with an exact recipe is that everyone follows it the same way. Give a thousand agents the same instructions for not looking generic and they end up looking like each other. And it's starting to show.
What the recipes ask for
All of this is written in the skill files, with these exact values:
- Four fonts instead of one. Inter is banned; use Geist, Outfit, Cabinet Grotesk or Satoshi. Swapping one overused font for four overused fonts only buys you time.
- The double-bezel card. An outer shell with a barely-there fill, a one-pixel ring, 6 or 8 px of padding and 32 px corners (
rounded-[2rem]), with a second card nested inside. One variant calls it “Doppelrand” and wants it on every major card. - The arrow in its own circle. If a button has an arrow, it sits inside a 32 px circle flush against the button's right edge. Never loose next to the label.
- Picsum photos.
picsum.photos/seed/…instead of Unsplash, sometimes in grayscale or withmix-blend-luminosityso they “don't look like stock”. - Deliberately messy numbers. To avoid 99.99%, the first version gives
47.2%and the phone number+1 (312) 847-1928as examples. Some agents paste the example as is. - A diffused shadow and 40 px corners on every large container:
shadow-[0_20px_40px_-15px_rgba(0,0,0,0.05)]androunded-[2.5rem].
None of these is wrong on one website. They're wrong on every website at once.
Why it happens
It's not the skill's fault. It does what it set out to do: get the AI away from Inter, purple and three cards with icons. But a model doesn't have taste, it has instructions. Replace one default with another fixed value and you've just made a new default.
When I calibrated the Slop Score, the thing that best separated generic sites from considered ones wasn't a color or a font. It was never having changed a single factory value. A good site doesn't stand out because of the font it uses but because someone chose it for that project. The guide on what AI slop is goes deeper.
How to use a skill without inheriting its fingerprint
- Keep the bans, drop the recipes. The bans work for everyone. The exact recipes are what repeats.
- Swap its examples for yours. If the skill says “Satoshi or Cabinet Grotesk”, write the font you picked into your brief and say why. Same for corner radius and colors.
- Search your code before you ship for
picsum.photos,rounded-[2rem],47.2%and-15px. If they're there, the recipe made it to production untouched. - Use real photos. A photo of your own, even an imperfect one, says more than a grayscale picsum shot.
For that brief, the prompts guide has one you can copy and adapt.
Frequently asked questions
What is taste-skill?
An MIT-licensed set of skills for coding agents that tries to keep the designs they generate from looking generic. The first version and its variants (soft, minimalist, brutalist and others) live in the taste-skill-classics repository.
Does using an anti-slop skill make my site look AI-generated?
Not by itself. Applying its examples without changing them does. The bans help; the fonts, radii and photos it recommends become a new fingerprint once they repeat across many sites.