Routing

Give advice without executing commands; the menu below is only for bare invocations. Consult relevant command references as needed for prerequisites and scope. Link to the docs (opens in a new tab) for the broader workflow guide. If the user also requests execution, follow that request.

No-argument routing: the context-aware menu

Read this when the user invokes /impeccable with no argument. They are asking “what should I do?” Make the menu context-aware instead of static.

Setup has already run impeccable context. If that reported NO_PRODUCT_MD, the project has no captured context yet: lead the menu with /impeccable init as the top recommendation (one line on why) and still show the rest below; don’t silently jump into init. Otherwise run "<skill-base-dir>/scripts/impeccable" signals once and read its JSON, then lead with the 2-3 highest-value next commands, each with a one-line reason pulled from the signals, followed by the full menu (the Commands table in SKILL.md, grouped by category). Never auto-run a command; the recommendation is a suggestion the user confirms.

Reason over the signals; there is no score to obey:

• setup.hasDesign false while setup.hasCode true → document (capture the visual system).
• critique.latest is null → the project has never been critiqued; for a set-up project with a real surface, offering /impeccable critique <surface> is a strong default.
• critique.latest with a low score or non-zero p0 / p1 → polish (it reads that snapshot as its backlog and closes it when stale or cleared).
• git.changedFiles pointing at one surface → scope audit or polish to those files specifically, naming them.
• devServer.running true → live is available for in-browser iteration, and generate for one-shot variant runs on a named element; if false, don’t lead with either. live, generate, and the bundled impeccable detect are web-only. If setup.platform is ios, android, or adaptive, don’t lead with any of them; the browser overlay and the HTML rule engine don’t apply to native app code.
• Otherwise group by intent (build new / improve what’s there / iterate visually), tailored to the current surface and setup.platform.

If scan.targets is non-empty and setup.platform is not ios/android/adaptive, run "<skill-base-dir>/scripts/impeccable" detect --json <scan.targets joined by spaces> once (the bundled detector over local files: no network, no npx; it reads HTML/CSS, so skip it for native projects). scan.via tells you what they are: git-changes (the markup/style files in your dirty tree, the most relevant set), source-dir (e.g. src, app), html, or root. Fold the hits into your picks: many quality / contrast hits → audit or polish; a specific slop family → the matching command (gradient text or eyebrows → quieter / typeset, flat or gray palette → colorize, and so on). It’s a real, current signal that beats guessing. If detect errors or the tree is large and slow, skip it and recommend the user run audit themselves; never block the suggestion on it.

Keep it to 2-3 pointed picks with the exact command to type. The menu stays the fallback; the recommendation is the lede.