A Claude Code skill that turns the final design polish of any UI/frontend build into a fixed, multi-agent amplification pipeline — and that first puts your entire installed design arsenal on the table instead of quietly using the same two skills every time.
npx skills add Shavy72/auto-polish-skill
- Good skills sit unused. You install great design skills and Claude keeps reaching for its favourites. auto-polish scans your whole setup — skills, MCP servers, plugins, AI CLIs — and puts every design-capable one to work.
- Claude never combines them. Left alone it applies one aesthetic lens, never several in service of a factorial jump. Here it is forced to, and you never have to remember which of your skills would have been the right one for this job.
Phase 0 — arsenal discovery + picker. A dependency-free Node scan reads every skill's frontmatter, your MCP config, plugins and PATH, classifies what is design-capable, and enriches each entry once with a concrete "specialised in …" line. The result opens as a local picker page where everything is pre-selected and you deselect what does not fit this job. Tools that need you in the loop (logins, paid steps) are flagged. Your selection returns through any of three channels: a local endpoint, a copy-paste token, or a downloaded JSON.
Phase 1 — 5x, parallel worlds. The selected tools are cut into 3–6 coherent bundles, and every bundle builds a real, runnable variant in its own worktree. Build green, desktop shot, mobile shot — and a video for anything animated, because screenshots prove nothing about motion.
Merge. A judge panel plus external AI tutors (Codex, Gemini — with images, not just code) pick the variant with the strongest structure as the base, then list every other world's strengths as transplants. Each is grafted in and made compatible with the base's material language, or rejected with a reason. The transplant table ships in the final report.
Phase 2 — 10x, condensed. No more fan-out: only the bundles that demonstrably won continue. Same improvement brief, applied harder. Then a minimalist judge hunts for what to remove.
Phase 3 — +3x. Final jumps plus every rollback from the removal list, then strict acceptance across desktop, mobile, motion and code.
Improvement gate. The factors are not arithmetic — they are the same improvement brief sent through the system repeatedly. A quota per stage plus a blind before/after "same league or a different league?" verdict stops alibi rounds; a missed gate gets one retry and is otherwise declared openly.
Born from a real production session where "make this update button 5x better, look at it, then 10x, then calculate one more 3x jump" produced dramatically better UI than a single polish pass — and where a missed button-label critique taught the skill that copy is a first-class design dimension.
npx skills add Shavy72/auto-polish-skill
Global install (available in every project):
npx skills add -g Shavy72/auto-polish-skill
/plugin marketplace add Shavy72/auto-polish-skill
/plugin install auto-polish@auto-polish-skill
git clone https://github.com/Shavy72/auto-polish-skill ~/.claude/skills/auto-polish
A manual copy needs SKILL.md, scripts/ and reference/ together:
~/.claude/skills/auto-polish/SKILL.md
~/.claude/skills/auto-polish/scripts/scan-arsenal.mjs
~/.claude/skills/auto-polish/scripts/picker.mjs
~/.claude/skills/auto-polish/reference/design-basis.md
To make it fire automatically, add a one-liner to your auto-active skills list in
~/.claude/CLAUDE.md, e.g.:
auto-polish — mandatory end phase after every new UI/frontend build: workflow 5x → judge → 10x → judge → +3x with resource inventory + desktop/mobile/video gates. Mini-addons (<~30 lines) on demand only.
- Claude Code with multi-agent orchestration (the
Workflow/Agenttools) - Node ≥18 for the two scripts — no npm install, no dependencies, works on Windows/macOS/Linux
- Playwright available via
npxfor screenshots/video (installed on demand) - Optional but recommended: external CLI models for the tutor round (OpenAI Codex CLI, Google Gemini CLI), a video-analysis MCP for motion review, a browser MCP for live acceptance
| Script | Purpose |
|---|---|
scripts/scan-arsenal.mjs |
Scans skills/MCPs/plugins/CLIs → inventory.json. --force rescans everything, results are mtime-cached. |
scripts/picker.mjs |
Renders the picker page and runs the local return channel → selection.json. --goal, --serve, --lang de|en, --port, --timeout. |
Both are safe to run standalone:
node scripts/scan-arsenal.mjs
node scripts/picker.mjs --goal "more modern and more premium" --serve
The server binds 127.0.0.1 only and accepts exactly two routes: GET / (the page) and
POST /save (your selection).
- Radical first, targeted second — build parallel worlds in practice, then merge the winners. Comparing real results beats reasoning about which approach should be best.
- Nothing gets forgotten — every selected tool lands in exactly one bundle, and the final report lists what went unused and why.
- Fixed stages, no open end — predictable cost, proven dramaturgy.
- Copy is design — button labels and microcopy get audited as hard as pixels.
- Done means done — no half stages; proof (build, screenshots, video) or it didn't happen.
This file is meant to be fine-tuned continuously in use. PRs and forks welcome.
MIT — see LICENSE.