Critique Skill

What is the Impeccable critique skill?

Critique is one of 24 commands inside Impeccable, a design skill for AI coding agents. You run it as /impeccable critique on a page, component or URL. It loads reference/critique.md and produces a written design review.

It does not edit your code. It scores the interface, names the problems in priority order, asks a few targeted questions, and ends with the commands to run next, finishing with polish.

What the critique report contains

• A design specificity verdict: is the look grounded in this product, or could any unrelated product reuse it unchanged?
• Nielsen's 10 usability heuristics, each scored 0 to 4, in a table with a specific finding per row.
• A cognitive load check against an 8-item list, plus any decision point with more than four visible options.
• Two to three strengths, three to five priority issues, persona red flags and minor observations.
• Recommended next commands, each tied to a finding, so the review turns into work.

How critique works, in order

1. Resolve the target and read .impeccable/critique/ignore.md, which silently drops findings you have already dismissed.
2. Run Assessment A, the design review, and Assessment B, the detector plus browser evidence, as two isolated sub-agents.
3. Merge both into one report with scores, ranked issues and personas.
4. Save a snapshot to .impeccable/critique/ and print the trend of your last 5 scores.
5. Ask two to four targeted questions, then list the recommended commands.

Who Impeccable critique is for

Best for

• A built page you suspect is generic or confusing but cannot say why.
• A review before launch with scored evidence in place of opinions.
• Tracking whether a screen improves from one pass to the next.
• Deciding which Impeccable command to run next instead of guessing.

Not for

• Making the fixes. Critique reports; typeset, layout and polish edit.
• A rating you can quote as fact. Scores come from a model and a rubric, and the guide says most real interfaces land between 20 and 32.
• Producing WordPress output. It reviews front-end work in your project.

Use it with UiChemy

Critique tells you what is wrong with a build. When the build is final and has to live in WordPress, UiChemy AI to WordPress imports an existing AI build as an editable page. Click Get Started on uichemy.com.

How to install Critique

npx impeccable install

1. Run the installer from your project root

   Run npx impeccable install. It lists the agent folders it detects, such as ~/.claude, ~/.codex or a project-local .cursor, lets you keep or change the set, then asks whether to install for this project or globally.

2. Skip the prompts in scripts

   Pass --providers=claude,codex,cursor and --scope=project (or global) to choose without questions. On Claude Code, Cursor, Codex, GitHub Copilot and Grok Build the installer also writes the provider-native design hook for the current project.

3. Reload your agent

   Restart or reload the harness so it picks up the new skill. In Codex, open /skills or type $impeccable.

4. Run init once

   Run /impeccable init. It reads the project, asks only for missing product facts, and writes PRODUCT.md so later commands know the audience, purpose and constraints.

5. Run critique

   Run /impeccable critique the checkout page. The target is optional and can be a page, a component or a file. /impeccable pin critique creates a standalone /critique shortcut.

Check that it loaded

Type /impeccable on its own. It shows the command list, and critique should be in it.

Read the official instructions on GitHub

Key features

• Two isolated assessmentsAssessment A is a design review and Assessment B is the detector scan with browser evidence. They must run as separate sub-agents whenever the harness has a sub-agent tool, so the detector cannot bias the design judgment.
• Degraded-run bannerIf sub-agents are unavailable or declined, the report's first line says so: DEGRADED, single-context, with the reason. A silent degraded critique is defined as a failed critique.
• Design specificity verdictThe agent judges whether composition, interaction and visual language are grounded in this product or interchangeable with any other, and makes that call before it sees detector output.
• Nielsen scoringAll 10 heuristics get a 0 to 4 score and a specific finding. Bands run from 36 to 40 Excellent down to 0 to 11 Critical, and the guide says most real interfaces score 20 to 32.
• Mode-aware n/aOn landing pages, campaigns and portfolios, heuristics 7 and 10 may be marked n/a. The total is renormalized, for example 24 out of 32, and the band is read from the percentage.
• Cognitive load checklistEight checks: single focus, chunking, grouping, visual hierarchy, one thing at a time, minimal choices, working memory and progressive disclosure. Zero or one failure is low load, four or more is high.
• P0 to P3 severityEach priority issue is tagged Blocking, Major, Minor or Polish, with why it matters, the fix and a suggested command. The test question: would a user contact support about this?
• Persona red flagsAlex the power user, Jordan the first-timer, Sam the accessibility-dependent user, Riley the stress tester and Casey the distracted mobile user. A landing page maps to Jordan, Riley and Casey.
• Detector and overlayimpeccable detect --json exits 0 when clean and 2 when it finds issues. With browser mutation available, a script injected into a tab labelled [Human] highlights the detected issues on the page.
• Snapshots and trendThe report is saved under .impeccable/critique/ with its score and maximum, then the last 5 scores print as a trend line. Polish reads the latest snapshot to pick up the priority issues.

Frequently asked questions