impeccable

Installation
SKILL.md

Ship production frontend. Refuse backend-only work. Done means the target is complete, responsive, on-brand, and visually verified.

Works the same in Claude Code and Codex. <skill-dir> below is the folder that holds this SKILL.md (for example ~/.claude/skills/impeccable, ~/.codex/skills/impeccable or .agents/skills/impeccable); use the real path in every node <skill-dir>/scripts/... command. Commands are written /impeccable <command>; in Codex type $impeccable <command>. "Ask the user" means your harness's question tool (AskUserQuestion in Claude Code, the user-input tool in Codex), or plain chat if there is none.

Run once per session: node <skill-dir>/scripts/context.mjs. Skip if its output is already in this conversation. On UPDATE_AVAILABLE, ignore it: this copy ships in AI Design PRO and an update would replace its edits.

No PRODUCT.md (NO_PRODUCT_MD): do not stop for an interview. If the page, brief, README or repo already says what the product is, who it is for and whether it is brand or product, infer the register from that and continue. Otherwise ask at most 2 short questions (what is it and for whom; brand or product surface), then continue. Write PRODUCT.md only if the user wants it. Run the full init.md flow only for init / teach or when the user asks for it.

Load references on demand, never upfront. If the first word is a command, read only reference/<command>.md. teach means init. Open another reference only when the one you are following links to it for a step you are actually doing, and read just that section. Do not read references "for context" or "to be thorough". Then read at least one existing CSS, token, theme, or component file.

Load brand.md when design is the product (marketing, landing, campaign, portfolio). Load product.md when design serves the product (app, admin, dashboard, tool). Load only one of the two, and only when the task sets or changes the look (craft, shape, bolder, quieter, colorize, typeset, layout, delight, overdrive, or no command on a redesign); skip both for audit, harden, adapt, optimize, clarify and extract. Pick by task cue, then the surface in focus, then register in PRODUCT.md.

If that file read found no committed brand colors, run node <skill-dir>/scripts/palette.mjs and compose OKLCH around the seed. Skip when existing tokens already own the palette. In shadcn projects, use theme tokens; never generic blue-* / sky-* / cyan-* / indigo-* accents.

Refuse: side-stripe borders, gradient text, decorative glass, hero-metric templates, identical icon-card grids, uppercase eyebrows on every section, numbered 01/02/03 scaffolding unless the section is a real sequence, overflowed headings, 1px border plus ≥16px blur shadow, card radius ≥24px, sketchy SVG, stripe backgrounds, "X theater" copy, em dashes, marketing buzzwords. Body contrast ≥4.5:1. Hero clamp() max ≤6rem. Display tracking ≥ -0.04em.

No argument: if init is already running, finish it. Else run node <skill-dir>/scripts/context-signals.mjs, recommend 2–3 commands, list the rest, and wait. Never auto-run. Prefer document when code exists without DESIGN.md; critique <surface> when never critiqued; polish when the latest critique has a low score or open p0/p1; scope audit / polish to dirty files on one surface; lead with live only if devServer.running. If scan.targets is set, run node <skill-dir>/scripts/detect.mjs --json <targets> and fold hits into the picks.

Commands: craft shape init document extract critique audit polish bolder quieter distill harden onboard animate colorize typeset layout delight overdrive clarify adapt optimize live. First word matching or clearly mapping to one: load its reference and treat the rest as the target. If two fit, ask once.

Installs
5
First Seen
4 days ago