What does this skill do, and when should you use it?
Impeccable is a frontend design skill for AI coding agents, derived from Anthropic's frontend-design skill and maintained by Paul Bakaus. It persists durable product truth in PRODUCT.md and visual direction in DESIGN.md via /impeccable init, then offers 24 sub-commands (audit, polish, bolder, animate, live, and more) covering review, refinement, motion, and in-browser live iteration. It ships 59 deterministic anti-pattern detector rules that run with no LLM and no API key, via a standalone CLI or a design hook. The repository is a monorepo bundling 24 skills; this profile covers only the skill at .claude/skills/impeccable/SKILL.md.
- Runs scripts/impeccable context to load PRODUCT.md, DESIGN.md, and the matching surface brief as persistent context for later commands
- Executes concrete operations via 24 commands: shape plans UX/UI before code, critique scores UX heuristically, audit checks a11y/performance/responsive, polish does a final pre-ship pass, harden adds error handling and i18n
- Provides refinement and enhancement commands: bolder, quieter, distill, typeset, layout, colorize, animate, delight, overdrive
- live and generate iterate visual variants on elements picked directly in the browser
- A design hook runs the detector after UI file edits and surfaces findings (on Claude Code, Copilot, Codex, Cursor, Grok Build)
- init, document, and extract write PRODUCT.md, generate DESIGN.md from existing code, and pull reusable tokens/components into a design system
- A frontend developer whose AI-generated landing pages always come out as purple gradients with nested cards uses audit and polish to systematically remove those template tells
- A product team planning a new dashboard uses shape and init to capture product context and design direction before writing code
- An indie developer finds their UI too bland and uses bolder, colorize, and delight to make it bolder — or too loud and uses quieter and distill to calm it
- A designer wants to iterate on hero-section variants against a real page in the browser using the live or generate command
- A team gates releases with harden (error states, i18n, edge cases) and runs npx impeccable detect in CI for deterministic checks
- Backend-only or non-UI tasks — the skill description explicitly excludes them
- Live variant editing on production sites — live mode only supports a local dev server or local static HTML; production sites are limited to read-only inspection via detect or the browser extension
- Environments that disallow binary downloads or hooks — the engine binary is fetched once to ~/.impeccable/bin/ on first run, which some team policies may block
How do you install this skill?
- First run downloads a closed-source engine binary into ~/.impeccable/bin/, statically unauditable; consider a controlled first run or building from source (cargo build -p impeccable).
- Install writes hook manifests into multiple harnesses, and hooks can trigger the download independently of session approval; review installed hooks before unattended runs (or pass --settings disableAllHooks).
- No Chinese-language support is disclosed and the engine binary is fetched from overseas servers; first-time installation may fail from mainland-China networks. A git submodule plus local build is a workaround.
- Live mode runs the target project's impeccable:manual-edit-validate script in a shell with your user permissions; use caution in untrusted checkouts.
- This is a static source review with nothing executed; the 59 detector rules and case-study effectiveness were not independently reproduced.
- Shell / CLI
- Network access
- Local filesystem
Self-contained engine binary (downloaded once to ~/.impeccable/bin/, no Node required except for npx installer)A local Chrome/Chromium/Edge for URL scans
The recommended route is the CLI installer, which detects installed AI tools and optionally installs the design hook:
npx impeccable installRefresh an existing install:
npx impeccable updateClaude Code (plugin marketplace):
/plugin marketplace add pbakaus/impeccableGitHub Copilot in VS Code (extension):
code --install-extension renaissance-geek.impeccableGrok Build (plugin):
grok plugin install pbakaus/impeccable#plugin --trustGit submodule (vendored teams):
git submodule add https://github.com/pbakaus/impeccable .impeccable
npx impeccable link --source=.impeccable --providers=claude,cursorAlso available: download a ZIP from impeccable.style or copy from dist/ into Cursor, Codex, Gemini CLI, Hermes, Pi, Trae, Rovo Dev, Qoder, Mistral Vibe, DeepSeek Harness, OpenCode, Google Antigravity, and others.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- /impeccable audit the landing page: check accessibility, performance, and responsive issues
- /impeccable critique the homepage hero section with a UX review
- /impeccable bolder this card is too timid, make it stand out
- /impeccable harden the checkout form: add error states, i18n, and edge cases
After install, everything routes through a single entry point, /impeccable. Start new projects with:
/impeccable initIt inspects the project, asks only for material gaps in durable product truth, and writes PRODUCT.md. Then invoke commands as needed:
/impeccable audit blog
/impeccable critique landing
/impeccable polish settings
/impeccable harden checkoutTyping /impeccable with no argument shows the command menu. Pin frequently used commands as standalone shortcuts (pin audit creates /audit). The hook auto-runs the detector after UI file edits; manage it with /impeccable hooks on|off. Use /impeccable doctor to diagnose and repair drift between project artifacts and this skill version. Codex users must approve the project hook in /hooks after install.
What are this skill's strengths and limitations?
- Very broad coverage: one command vocabulary spans planning (shape), review (critique/audit), refinement, motion, and in-browser live iteration
- 59 deterministic detector rules run with no LLM and no API key; the CLI can run standalone in CI
- Explicit anti-pattern guidance (overused Inter font, purple gradients, nested cards, gray text on color) directly targets the common AI-slop look
- Persistent PRODUCT.md and DESIGN.md mean later commands know audience, purpose, and visual world instead of starting from scratch
- No Node or other runtime needed after install — the engine is a self-contained binary
- The engine binary downloads over the network on first run, and hooks can trigger this download even in unattended runs — the README itself flags reviewing hook security
- Live mode does not support production sites; real-time editing is limited to trusted local projects
- Command quality depends on how well the host AI tool follows the skill; the source provides no independent effectiveness benchmarks
- Hook behavior varies widely across tools (no hook on Hermes, no edit hooks on Veto), so the experience is not uniform
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Impeccable this page | 58 · Recommended | ★ 79k | 2d ago | Apache-2.0 |
| TypeUI Design Fundamentals | 70 · Strongly recommended | ★ 2k | 3mo ago | NOASSERTION |
| Distinctive Frontend Design ✓ Anthropic · Official | 60 · Recommended | ★ 38k | 3d ago | Apache-2.0 |
| SwiftUI Design Skill | 57 · Use with care | ★ 211 | 5mo ago | MIT |
| Frontend Design Skill | 52 · Use with care | ★ 5.5k | 5mo ago | MIT |
The README explicitly states the project started from Anthropic's official frontend-design skill and builds on it with an init flow, 24 commands, deterministic detector rules, live browser iteration, and multi-tool installation support.
How did FollowSkills review this skill?
Evidence shows restrained scope: hooks limited to UI-edit detection, hooks on/off/ignore controls, confirmation required for replacing factual copy or repairing drift, explicit refusal to inject into production sites, and prompts to review installed hooks. However, first run downloads a closed-source engine binary into ~/.impeccable/bin/, statically unauditable; hooks can trigger the download independently of session approval. Deducted for lack of hash/signature verification evidence and incomplete data-flow disclosure around the binary.
Docs are highly self-consistent: explicit launcher-failure fallback, CONTEXT_STALE handling, doctor repair, Windows .cmd branch, and bounded verification rounds preventing runaway loops. But key paths depend on an unreadable prebuilt binary; script/binary consistency cannot be confirmed statically. Failure feedback is well described but unverifiable. Capped at 10 per static calibration.
Trigger definition is excellent: description enumerates fit and explicit non-fit ('Not for backend-only'), full command table, routing rules, and menu-only behavior on bare invocation — supporting accurate semantic invocation. Deducted for: no Chinese-language support disclosed, and core function depends on a first-run engine binary downloaded from overseas servers, a mainland-China reachability risk with no mirror provided.
Version (4.5.2), Apache-2.0 license, progressive-disclosure architecture (SKILL.md → layered reference/), and known-limitation disclosure (no production-site live mode, DevTools limits) are all complete, with clear repo-level update paths (npx update, submodule). Deducted: no changelog within the skill itself, and the binary engine version (Cargo 0.1.14) is not mapped to the skill version inside the skill.
Outputs are well specified (audit reports, modified UI code), command structure is complete with clear boundaries, and marginal value over manual design review is plausible with an external case-study link. But output quality cannot be verified statically and the core detector is unreadable; capped at 7 per static calibration.
Substantial auditable primary material is present (layered reference docs, Cargo workspace source, claimed 59 deterministic rules, gitignore manifests) consistent with the claims. However no committed test suite or CI evidence covering the skill's key paths appears in the provided files, and the 59 rules and case study cannot be independently reproduced in this review. Scored 4 within the static cap; deducted for missing executable verification evidence.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →