Impeccable Skill Review: Verdict, Safety and Alternatives

The Impeccable skill is Paul Bakaus' design bundle: one router skill with 24 commands, an engine and hooks. Our verdict is Conditional, 67/100, with safety grade C. We reviewed skill 4.5.0 on 2026-09-30, weighing 20 dated reports, 9 of them strong. Four releases have shipped since then, and the verdict below covers 4.5.0 only.

Get the skill: impeccable folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing

What is the Impeccable skill, and what does it do?

The Impeccable skill is Paul Bakaus' frontend design bundle for coding agents. It holds one router skill with 24 commands, such as critique, audit and polish. A Rust engine runs an anti-pattern detector through edit hooks.

Impeccable is an agent skill bundle for frontend work. Its job is to make a coding agent produce, critique and refine interfaces that look deliberately designed. It steers the agent away from template output. The router SKILL.md sends each command to its own reference files, so most guidance loads per task. The guidance is split by surface mode, which means brand pages and product screens get different rules. Our notes on how SKILL.md files are structured explain the router pattern in general terms.

Beyond Markdown, the bundle ships a launcher, a Rust engine, plugin hooks, a live browser mode and helper agents. Hooks call the engine after every edit, and its detector flags AI-made design patterns.

Which Impeccable sites and versions are official?

The official Impeccable sources are four: impeccable.style, its docs, the pbakaus/impeccable repo and the impeccablestyle Substack. Many third-party guides still describe versions 3.0 and 4.0. We reviewed version 4.5.0, and 4.7.0 is current.

Command and rule counts change with each version, and the dev.to guide lists 23 commands and 58 rules. The 4.5.0 SKILL.md we reviewed lists 24 commands. The site's 61 checks are the publisher's own figure, dated 2026-10-21.

Is Impeccable a good skill to install?

Yes, Impeccable is a good skill to install for critique and polish on macOS or Linux. It scores 67/100, Conditional, at Moderate confidence. Its guidance is specific, but an unsigned engine runs on every edit.

Impeccable pairs strong, specific critique and polish guidance and a fast-moving maintainer with a heavy runtime. That runtime brings an unsigned downloaded engine, hooks on every edit, Windows breakage and a risky live mode. The strength is concrete, opinionated guidance with a critique score that reviewers found accurate. The maintainer fixed reported bugs such as #400 and #304 within days.

The weakness sits in delivery. Much of the value lives behind a launcher and 85K tokens of references. Agents often never load them, per issue #744. Live-wrap wrote invalid TSX (#852), Windows hooks failed (#878), and the engine is unsigned (#740).

Impeccable suits developers on macOS or Linux using Claude Code, Codex or Cursor. They want a structured critique, audit and polish pass on an existing web interface. They accept a binary engine plus edit hooks. It is a poor fit in two cases:

• Windows users and locked-down machines, where hooks failed under PowerShell and antivirus tools flag the unsigned engine.
• Teams that forbid a skill from downloading a binary, pinging a remote roll service or spawning permission-bypassing agents.

The design guidance alone scores 78. The worst failure is live mode silently corrupting source (#852). That failure and the hook-plus-binary runtime pull the overall score to 67. The problems concentrate on Windows and live mode, so the score sits in the Conditional band. Compare it with the best UI/UX design skills for Claude before you choose.

No measured result exists yet: only anecdotal reviews and the publisher's earlier figures from Tessl, a skill registry. Our page on how verdicts are scored sets out the bands and caps.

Two pieces of evidence are missing. Nobody has published a controlled with-and-without run on current models. Such a run measures design quality, reference loading and token cost. A live-mode test on multi-line JSX after PR #916 is missing too.

Test it yourself:

• Install Impeccable on a throwaway repo on macOS or Linux.
• Run /impeccable critique on one existing page.
• Check whether the agent reads reference/critique.md before editing.
• Check whether it reads craft-floor.md before editing.

How many tokens does Impeccable use?

Impeccable's SKILL.md measures about 3,029 tokens per load. Each command pulls an estimated 15,000 to 25,000 tokens of references. Our estimate for polishing one landing page is 40,000 input and 10,000 output tokens.

The engine, tens of MB, downloads to ~/.impeccable on first run. Hooks run after every Edit or Write, and a deep pass runs on Stop. Each finding adds latency and agent turns. Optional image generation bills your OpenAI key.

What changed since our review?

Four Impeccable releases followed our review, ending at 4.7.0 on 2026-10-19. The publisher reports fixes for Windows hooks, live-mode wrapping and telemetry opt-out. We have not re-tested them, so the verdict stands.

Issue states moved too, as of 2026-10-21. Fix PR #916 for #852 merged on 2026-10-04, while #852 itself stays open. Issues #878, #884 and #857 closed on 2026-10-15, and #853 closed on 2026-10-16. Issue #813 on engine integrity and #744 on reference adherence remain open. No release note mentions Windows code signing.

Is the Impeccable skill safe to install?

Yes, the Impeccable skill is safe to install where a downloaded binary is acceptable. Its safety grade is C. Hooks run an unsigned engine after every edit, and the engine calls impeccable.style unless you opt out.

Grade C means the skill needs credentials or broad permissions. It describes access, not intent, and it covers the files at the pinned commit only. Our guide to vetting skills that run hooks and binaries explains each check. What we found in the files at 4.5.0:

1. SKILL.md is 12,115 bytes, about 3K tokens. It routes to 45 reference files totalling about 85.6K tokens, loaded per command.
2. Two launchers, scripts/impeccable and impeccable.cmd, run a sibling or cached binary. Otherwise they download engine-v<version> from GitHub releases.
3. The launchers verify a same-release .sha256 sidecar and fail closed.
4. The plugin hooks.json wires SessionStart, PostToolUse (Edit|Write) and Stop to the engine. The engine therefore runs after every edit, and the guard is POSIX-only.
5. The engine checks impeccable.style/api/version daily. Concept-seed calls roll endpoints with an anonymous choice ping.
6. Three variables opt out: IMPECCABLE_NO_TELEMETRY, DO_NOT_TRACK and IMPECCABLE_NO_UPDATE_CHECK. Run output and docs disclosed the calls; the README did not.
7. Live copy-edit apply defaults to auto. It can run codex exec or claude with permission bypass on page-derived prompts, hardened after #507.
8. Live browser scripts talk only to a localhost server with a token. The wildcard CORS token leak (#304) was fixed in v4.0.2.
9. OPENAI_API_KEY, when present, enables image generation for comps. No other credentials are required.
10. The bundle writes PRODUCT.md, DESIGN.md, .impeccable/ config, mocks and surface briefs. Caches go under ~/.impeccable.

The Windows engine is not code-signed, and ESET and Defender flagged it (#740, #814). The skills.sh directory of Agent Skills shows three audits: Gen Agent Trust Hub Pass, Snyk Pass and Socket Warn. The checksum comes from the same release as the binary, so it does not prove origin. Issue #813 tracks that gap and is still open.

The official sources disagree on two points. The impeccable.style homepage does not mention telemetry, while the engine output and docs disclose it. The homepage describes no engine download, while we found the launcher fetching one from GitHub releases. Release 4.6.0 adds impeccable telemetry off, per its notes.

What does the community report about Impeccable?

The community reports about Impeccable describe useful critique and fast maintainer fixes. One site rose from 25/40 to 32/40 on its findings. Issue #744 found agents skipping every reference file, and HN users report converging output.

Community evidence comes from GitHub issues, Hacker News and independent reviews dated before 2026-09-30. The measured run in #744 and the reproduced #852 explain the reliability score of 52.

Negative items cluster on runtime and delivery, and positive items on the design advice.

What are the Impeccable skill's known failures?

At skill 4.5.0, the Impeccable skill had six known failures. Our reliability score counts four: invalid TSX from live-wrap (#852) and Windows hook failures (#878). The others are detector gaps on framework files (#884) and skipped references (#744).

After-review statuses come from release notes or GitHub issue states on 2026-10-21.

1. Live-wrap source corruption (#852). Multi-line self-closing JSX overshot by up to 50 lines and deleted sibling components. At 4.5.0: reproduced by the maintainer. After review: fix PR #916 merged on 2026-10-04, and the issue is still open.
2. Windows hooks (#878). All three Claude Code plugin hooks failed under PowerShell. At 4.5.0: confirmed by the maintainer. After review: closed on 2026-10-15, with PowerShell hooks reported in 4.5.1.
3. Detector gaps (#884). Framework files caught fewer rules than byte-identical HTML. At 4.5.0: open. After review: closed on 2026-10-15.
4. Skipped references (#744). On 4.2.1 with Haiku 4.5, agents read 0 of 35 references and refused 55 of 55 launcher calls. At 4.5.0: a 4.2.2 fallback existed, which the maintainer calls partial. After review: still open.
5. Installer outage (#857). Installs failed with HTTP 404 for days. At 4.5.0: open. After review: closed on 2026-10-15.
6. Detector false positives (#881, #882). Users reported contrast and occlusion false positives. At 4.5.0: open. After review: releases 4.5.1 and 4.7.0 report fewer of them.

Skipped references are a general agent behaviour, and our guide on why skills skip their references covers the causes. The worst failure gives no warning. Live mode rewrites source into invalid code while the UI looks fine. A hook-driven binary misbehaving on every edit is the second form.

Does Impeccable work in Codex, Cursor and other agents?

Yes, Impeccable works in Claude Code and Codex, where it is verified. It ships provider folders for Cursor, Gemini CLI, GitHub Copilot, OpenCode and other agents, which remain claimed. Windows was its weak platform at review.

Claimed means the publisher lists the agent without evidence confirming it. Impeccable needs Bash or PowerShell, curl or wget for the first engine download, and browser automation for live mode. Windows shells broke hooks in three reports (#848, #859, #878). Our list of Claude Code skills ranked places Impeccable among other skills for that agent.

How do you install the Impeccable skill?

You can install the Impeccable skill with npx impeccable install, which needs Node.js 18.19 or later. In Claude Code, you can add the pbakaus/impeccable plugin marketplace instead. The engine downloads on first run.

The installer needs Node.js, per the tutorial, while the engine itself is a binary. Launcher comments say the engine needs no Node, a publisher claim.

In Claude Code

Pick one of three install paths:

npx impeccable install

/plugin marketplace add pbakaus/impeccable

npx skills add pbakaus/impeccable

After the marketplace command, open /plugin and install Impeccable from the list. The impeccable.style homepage shows the npx skills add form. Then turn on the hooks so the detector checks each edit:

/impeccable hooks on

Approve the hooks only on a machine where a downloaded binary is acceptable. The general steps for every skill are in our guide to Claude Code skill installation.

In Codex, Cursor and other agents

npx impeccable install sets up each supported harness from one command. Teams can vendor the repo as a git submodule and link it per provider:

npx impeccable link --source=vendor/impeccable --providers=claude,cursor

The tutorial says GitHub Copilot users enable skills via Settings, Experimental. No published test covers install paths outside Claude Code and Codex.

How do you use the Impeccable skill in Claude Code?

You use the Impeccable skill in Claude Code by running /impeccable init once per project. Then call a command such as critique, audit or polish. Approve the hooks so the detector checks each edit.

A first pass on an existing page takes seven steps:

1. Run /impeccable init once per project.
2. Describe your users and product in PRODUCT.md.
3. Run /impeccable document to write DESIGN.md, an optional step.
4. Pick the task: a new build or an improvement to an existing page.
5. Run /impeccable critique for a UX review, or /impeccable audit for technical checks.
6. Run /impeccable polish to apply the last-mile fixes.
7. Review the findings and the diff before you accept anything.

The 4.5.0 SKILL.md sorts its 24 commands into six categories:

A standalone polish skill is also listed on skills.sh. At review it had 87.4K installs, critique 84.8K and audit 84.0K. In 4.5.0 these run as /impeccable polish and its siblings. The impeccable pin command restores a standalone /polish shortcut.

The evidence supports two primary tasks:

1. Critique or audit an existing site and fix the findings, weekly (mejba.me, react-japan.dev).
2. Polish a working UI with typeset, layout, polish and distill, weekly (react-japan.dev, #149).

Three secondary uses appear in the evidence. Developers give a new landing page a distinct visual direction now and then (#744, HN 49059235). Others run the detector as a lint step (HN 49294522, #884). Some iterate element variants in live mode (#852, mejba.me). Documenting an existing system as DESIGN.md is claimed only, because #744 found the skill not activated for it.

What are the alternatives to the Impeccable skill?

The closest alternatives to the Impeccable skill are Frontend Design and UI/UX Pro Max. Frontend Design scores 76, Recommended, for new pages. UI/UX Pro Max scores 68 and Taste Skill 62, both Conditional.

None of the four needs a downloaded engine. The page with all rated design skills holds the full field.

Where is the Impeccable skill on GitHub?

The Impeccable skill is on GitHub in the pbakaus/impeccable repository, with the skill at .claude/skills/impeccable. We reviewed tag skill-v4.5.0, commit 508d7e8. The repo shipped 47 releases in the 90 days before our review.

• Repository: github.com/pbakaus/impeccable, Apache-2.0.
• Skill path: .claude/skills/impeccable, 57 files.
• Reviewed commit: 508d7e8, identical to the skill-v4.5.0 tag published 2026-10-02.
• Activity at review: 50 contributors, 47 releases in 90 days and 77 in 12 months.
• Pull requests in the period we reviewed: 398 merged and 104 unmerged.
• Distribution: a Claude Code plugin marketplace plus provider folders for other agents.

Common questions

Is Impeccable free?

Yes. Impeccable is free under Apache-2.0, and only optional image generation bills an OpenAI key.

Does Impeccable work on Windows?

Partly. At 4.5.0 its hooks failed under PowerShell (#878), and antivirus tools flagged the unsigned engine (#740).

Can I turn off Impeccable's telemetry?

Yes. Set IMPECCABLE_NO_TELEMETRY or DO_NOT_TRACK, and 4.6.0 adds impeccable telemetry off (per its release notes).

Do I need the Higgsfield MCP for Impeccable?

No. The chaseai.io guide names it, but we found no required MCP server.

How do you write a skill like Impeccable?

Start with one SKILL.md and add reference files only when a command needs them. Our guide on how to write your own skill walks through each step.