Unslopping AI output with nothing but a skill

You do not need a better model to get rid of the purple gradients. You need your taste written down as a skill the agent loads before it touches the screen.

You can spot a vibecoded screen from across the room. Purple gradient, glowing cards, an emoji on every label, a headline that shouts. The model is not broken. Nobody told it what you consider good, so it gave you the average of everything it has seen.

The usual fixes are a longer prompt, a bigger model or a plugin. I think the cheapest one is a skill.

What a skill is

A skill is a folder with a SKILL.md inside. The agent sees a one-line description up front and loads the full instructions only when the task calls for them. There is no fine-tune, no server and no API. It is taste and procedure written in plain text, sitting in your repo.

What unslop gets right

unslop by Damir Sokolovsky is the best example I have read. Three things make it work.

It audits before it edits. The skill ships a list of signs, each paired with its fix. Gradient text becomes a solid colour. An arbitrary mt-[13px] moves onto a 4px scale. The agent has to name what is wrong, and where, before it changes a line.

It fixes in a fixed order. Spacing, colour, icons, copy, type, layout, details, visuals, motion. One layer across the whole screen, then the next, the way a designer does a redline review. Order matters because type choices mean little while the spacing is still random.

It knows when to stop. If a screen is already good, the skill says so and suggests two or three refinements. Most prompts cannot do that. Ask a model to improve something and it will always find something to change.

Writing your own

The same shape works well outside UI. I use it for specs, commit messages and code review.

Name the signs. Write the tells you actually reject, with the fix beside each. “Be concise” does nothing. “A summary that repeats the title, delete it” does.
Order the passes. Decide what gets fixed first and say why.
Define done. Give the skill permission to change nothing.
Keep the references next to it. Examples of good output teach more than adjectives.

Then put it in version control. A skill is reviewed, diffed and improved like any other code, and every agent on the team picks up the same standard from the next run onward.

Slop was never a model problem. It is what you get when the standard lives only in your head.