unslop claude skill: Fixing AI-Looking UI Goes Viral

The unslop claude skill is the dev-Twitter toy of the week. Created by designer Damir Sokolovsky, it promises to turn "vibecoded," AI-looking web interfaces into screens that look designed by a person. The GitHub repo, created on October 29, had about 5,100 stars at our check on November 2 and ranked second on Daily Launch's November 1 list.

What actually happened

AI coding agents made it trivial to ship a working dashboard in an afternoon. The side effect is that many of those dashboards look the same: gradient hero, glowing cards, emoji bullets, generic copy. Designers started calling it slop, and the unslop claude skill packages a fix.

According to the README, the skill first lists every sign of slop and where it lives. It then proposes a direction based on your product, and only after that edits the screen "pass by pass like a designer doing a redline review." If the screen is already good, it says so and suggests small tweaks instead of a redesign.

It is framework-agnostic. The README lists Next.js, React, Vue, Svelte, Tailwind and plain HTML and CSS. It can even work on a live site when all you have is a link, using Playwright to snapshot the page.

How the unslop claude skill works

Installation is a copy job. For Claude Code, you clone the repo and copy the skill folder into ~/.claude/skills/, or into a single project's .claude/skills/ folder. For Codex, the README says you can ask the agent to install it straight from the repo.

Usage is conversational:

• /unslop this dashboard
• /unslop the hero of a live URL
• /unslop my landing page, plus an optional "story video"

The story video feature records the before-and-after as a short clip for social media. That requires Playwright with Chromium plus ffmpeg on your machine.

The pitch is not "make it pretty." It is "make it look like a human with taste made a decision."

Why it went viral

Daily Launch, which ranks new projects every morning, placed unslop second on November 1 with about 2,200 stars in a day. Only DeepSeek's agent prototype ranked higher, which we covered in deepseek superagent prototype. GitHub's API showed about 5,100 stars and 255 forks when we checked on November 2.

The timing fits. Claude models keep getting better at code, as we saw with claude sonnet 5 5, and more non-designers are shipping front ends. A skill that encodes design taste into a reusable prompt pack hits a real pain point.

It also shows how the "skills" format is becoming a distribution channel. A skill is basically a folder of instructions and references. If it is good, people share it like a library.

Limits worth knowing

Start with expectations. The skill can clean up spacing, color and copy, but it cannot invent a brand you never defined. If your product has no clear audience or voice, the proposed direction will be a guess. The README itself frames the first step as setting a direction from your product, which only works if you can describe that product clearly.

A skill is not magic. It steers the model with instructions and examples, and the output still depends on the underlying model and your codebase. Results on a messy legacy project will vary.

There are no independent benchmarks here. The before-and-after clips in the repo are the author's showcase, not a controlled test. Treat them as a demo, not a guarantee.

The license matters too. CC BY 4.0 lets you use and adapt the work with attribution, which is unusual for code. Check whether that fits your company's policy before adding it to a commercial repo.

What to do as a reader

If you build with Claude Code or Codex, try the unslop claude skill on a throwaway branch first. Let it run the audit, read the proposed direction, and only accept the edits you agree with.

Read the unslop GitHub README for install steps and requirements before cloning. Keep your own design system as the source of truth, and use the skill as a second opinion, not the final word.