Impeccable
A design skill that teaches your agent to avoid the "AI-made" look and gives you commands to critique, audit and polish your interface.
The verdict
Only if you need it
Useful if how your app looks matters and you'll spend tokens on design passes. Its critique and audit commands get the warmest reviews. But it has grown into a big, fast-changing system with hooks, a downloaded binary and telemetry, so start small and don't expect it to make you a designer.
Best forPeople building front ends with an agent who keep getting the same purple-gradient, card-in-a-card look and want a second opinion on their UI.
- Stars
- 85.1k
- Licence
- Apache-2.0
- Language
- Rust
- Last release
- 19 Oct 2026
- Created
- 16 Nov 2025
- Open issues
- 30
As of 23 Oct 2026. Numbers move fast on trending repos.
npx impeccable installRun it from your project folder, then type /impeccable init inside Claude Code. In Claude Code you can instead run /plugin marketplace add pbakaus/impeccable and install it from the /plugin menu. There's also a VS Code extension for Copilot, and ZIP downloads for other tools at impeccable.style.
What it is
Ask any AI model to build a web page and you tend to get the same look: the Inter font everywhere, purple-to-blue gradients, cards inside cards, grey text on coloured backgrounds. Impeccable is a "skill", a set of written instructions your agent loads when it does design work, meant to steer it away from those habits.
It was created by Paul Bakaus, who built jQuery UI and worked as a developer advocate at Google. It started as an extension of Anthropic's own frontend-design skill and has grown a lot since. Today it's one skill with 24 commands, a design checker with 62 rules, a live mode for trying variations in your browser, and hooks that check your UI files automatically. It's free and Apache-licensed.
The project is very active: more than 20 releases since July 2026, and since September its checks run on a compiled engine written in Rust.
What actually happens when you use it
- Setup.
/impeccable initlooks at your project, asks a few questions about your product and audience, and writes aPRODUCT.mdfile so later commands know who the app is for. - Commands. Everything runs through
/impeccable <command>.critiquegives a design review.auditchecks accessibility, speed and mobile layout.polishdoes a final pass. Others likebolder,quieter,typesetandlayoutnudge one aspect of the design. - New designs. For a new page, it deals out a few visual directions from a catalogue on its website, and you or the agent pick one.
- Automatic checks. In Claude Code, Cursor, Codex and a few others, a hook runs its checker every time the agent edits a UI file and reports problems back to the agent.
- A standalone checker.
npx impeccable detect src/scans your files, or a live URL, for the same AI-made tells without using any AI at all.
What people like
- The critique is a good second opinion. One developer ran it on his portfolio, got 25 out of 40 on the first critique, fixed the two biggest problems and got 32. His lesson: fixes stick when made to your design tokens (shared colours, fonts and spacing), not page by page.
- It helps if you have no design background. A Hacker News commenter with zero design experience said its critique of their project felt useful.
- It keeps a design consistent. One user who compared it with other design skills on the same prompt liked that it writes documents to keep the design coherent, and found its results decent but with some AI-made touches left.
- Better than the built-in options for some. Another user said it still needs guiding but is much better than Claude's own design tool for them.
What people run into
- The launch examples looked worse, not better. At its January launch on Hacker News, many commenters said the "after" designs were worse than the "before" ones. Bakaus agreed they were rushed and took them down.
- Results vary. One user found it produced only sub-mediocre results. In a comparison thread, the tester slightly preferred a rival "Taste" skill for look and feel. Another posted a plain login form they said cost 150,000 tokens.
- It's heavy. The hands-on reviewer found the full command set takes a big share of the context window, so he only loads it for design sessions. An October issue says some commands run their most expensive process even for small requests. The maintainer agreed to trim some of that and kept other parts on purpose.
- It doesn't always follow its own process. An independent tester ran hundreds of trials and found the skill often switched on but never read its own reference guides. The maintainer called the first fix partial, and the issue is still open.
- Live mode is rough. The same reviewer called browser live mode alpha quality: sessions needing restarts, unreliable element picking, and one change written to the wrong file.
Before you install
- It downloads and runs a program. Since September, the skill runs a compiled engine that downloads on first use into
~/.impeccable/bin/. In Claude Code the hook can trigger that download on the first UI edit, even if you said no to the agent running it. In September, Windows antivirus tools flagged the then-unsigned engine. It's now signed. An open issue says the download is only checked against a file from the same place, and a fix is planned. - Telemetry is on by default. Starting a new design always asks impeccable.style for visual directions. Unless you opt out, it also records which option won. It doesn't send your code or files. Run
npx impeccable telemetry offor setDO_NOT_TRACK=1to switch the recording off. - Live mode runs code on your machine. It can run a script from your project's
package.json, so only use it on projects you trust. An earlier version leaked live mode's access token to other pages, which has since been fixed. - Expect frequent updates. New releases arrive almost weekly. Run
npx impeccable updatenow and then, and add the.gitignoreblock from the README so its working files stay out of your commits. - Try the checker first.
npx impeccable detectgives you most of the "spot the AI look" value with no AI, no tokens and no hooks.
The short version
Impeccable gives your agent a design vocabulary and a reviewer, and people who care about UI often find the critique and audit commands worth it. It won't turn generic output into great design by itself, it uses a lot of tokens, and it has grown into a complex system that changes every week. Start with the standalone checker or a single critique, and add the rest only if it earns its place.
Sources
GitHub
- The repo and README
- Skill 4.7.0 release
- Issue #744: the skill switches on but skips its own guides
- Issue #985: references demand full-cost work on small requests
- Issue #740: Windows antivirus flagged the unsigned engine
- Issue #813: engine download has no independent integrity check
- Issue #304: live mode leaked its access token (fixed)
Hacker News
The week’s repos, every Friday.
From the same drop
obra/superpowers
Superpowers
A set of skills that makes your coding agent brainstorm, plan, write tests and review its own work before it calls anything done.
★ 304kBreakdown →
deepseek-ai/deepseek-harness
DeepSeek Harness
DeepSeek's own open-source coding agent, run from a web page or desktop app, built so that almost every feature is a plugin you can swap.
★ 259kBreakdown →
anomalyco/opencode
OpenCode
The most popular open-source coding agent. It works like Claude Code, in the terminal, a desktop app or the browser, but with almost any AI model you like.
★ 220kBreakdown →