Source: impeccable

What it is

  • A large design skill for AI coding agents, by Paul Bakaus, one of jQuery UI's early creators.
  • Licence: Apache 2.0. It started from Anthropic's frontend-design skill. Its iOS and Android guidance comes from ehmo's MIT-licensed platform-design-skills.
  • Version: its npm page shows the current one.
  • Packaging:
    • one skill, impeccable, with a SKILL.md and about 40 reference playbooks;
    • four helper subagents (asset producer, documenter, finish reviewer, manual-edit applier);
    • a scripts/impeccable launcher that runs a self-contained Rust engine, shipped alongside or downloaded once;
    • ready-built copies for about 19 agent tools, plus a Claude Code plugin, a VS Code extension and a Chrome extension.
  • 24 commands, including craft, which is an old alias. Admin commands: pin/unpin, hooks, doctor, and teach (another name for init).
  • A rule checker with 62 built-in rules. No AI model is involved in checking.

How to get it

  • Install with npx impeccable install.
  • Reload your coding tool, then run /impeccable init once per project ($impeccable init on Codex).
  • Commands take the form /impeccable <command> <target>.
  • /impeccable doctor finds stale context files, unused settings and broken hooks.

Cautions

  • Its design hook runs unsandboxed, with the agent's environment and network access.
  • The engine may download a binary the first time it runs.
  • A clean rule-checker scan is a floor, not proof of quality. impeccable says so itself.
  • Some rules are given only for Codex runs (tracking, card corners, illustration, backgrounds, elevation). The drawers mark them.
  • Several claims are stated without evidence. The drawers mark these "(asserted, no evidence shown)". They include:
    • the before-and-after example in the README is one run, with no method shown;
    • the tells "show up again and again, across models and projects";
    • the hook is "where most slop gets caught";
    • AI-made interfaces "cluster around a few looks regardless of subject";
    • models "systematically believe" an image copy worked when it didn't;
    • side stripes, purple gradients and cyan on dark are "the most recognizable" tells;
    • "most real interfaces score 20-32 out of 40";
    • default live mode is "right for ~90% of sessions";
    • "German is often 30% longer than English";
    • 80 to 85% image quality is "usually imperceptible";
    • the four-item memory limit names a 2001 study but doesn't show it;
    • usage data is "how the catalog learns what works", with no mechanism shown.

Cards it supplies

  • Tool: design review with scores (impeccable critique).
  • Tool: technical audit (impeccable audit).
  • Tool: final quality pass (impeccable polish).
  • Tool: rule checker for AI tells (impeccable detect).
  • Tool: DESIGN.md, the look written down (shared with awesome-design-md).
  • Its other commands (shape, typeset, colorize, layout, bolder, quieter, distill, animate, harden, clarify, onboard, adapt, optimize) are named in the drawer where each one applies.

Leads onward

Leads here