Lyu Pro Skill Library

A curated library of hybrid skills for Claude / SkillForge.

A hybrid skill pairs a lightweight SKILL.md (routing + a short instruction injected into the prompt) with deterministic scripts (.mjs) the agent runs on demand. The deterministic part costs 0 prompt tokens, is reproducible, and is CI-ready.

The library is owned end-to-end — rules are tuned to our own stacks (React Native Web, web), our brand, and our anti-slop discipline. One library serves every SkillForge consumer through folder-registration.

Why hybrid

• 0-token determinism. Detectors and generators run as scripts; the model pays only for the light SKILL.md and the script's JSON result, never for hundreds of lines of rules.
• Reproducible & CI-ready. Deterministic scripts run the same in an agent prompt, a pre-commit hook, or a CI gate.
• Full control. No vendoring, no upstream drift, no foreign installer.

Layout

skills/
  design/   # design phase: design-taste, html-mockup-generator, style-*, brand-asset-protocol, design-system-advisor
  uiux/     # UI-UX audit: design-slop-detector, design-critique, a11y-detector
  meta/     # the library about itself: skill-lint
  git/      # (future)
  content/  # (future)
docs/
  skill-format.md      # the hybrid skill format (study before adding a skill)
  install-runbook.md   # SkillForge folder-registration + reindex

Categorization is physical, by folder — a design skill lives in skills/design/. See docs/skill-format.md.

Skills

| Skill | Category | Type | What it does | |---|---|---|---| | design-slop-detector | uiux | deterministic | 43 rules (39 always-on + 4 opt-in --tells) flag AI design-slop in HTML/CSS → JSON. The 0-token gate. | | a11y-detector | uiux | deterministic | 8 static accessibility rules (missing alt, unlabelled controls, empty buttons/links, duplicate ids, zoom locks, …) → JSON | | design-token-adherence | uiux | deterministic | 3 rules flag declared color / spacing / font-size values off your design-system token scale (caller supplies --tokens) → JSON. The system-consistency gate. | | ui-gate | uiux | deterministic | runs both detectors in one pass, merges findings (tagged by tool) → one verdict + one --fail-on gate. The single UI check for agents/CI. | | render-verification | uiux | deterministic* | renders HTML in headless Chromium and flags computed-layout bugs static analysis can't: horizontal overflow, text flush to the edge, real clipping, sub-min touch targets, off-screen content, screenshot diff → JSON. *Opt-in: needs the optional Playwright peer dependency. | | design-critique | uiux | prompt | evaluative pass: visual hierarchy / IA / cognitive load → score + P0–P3 | | design-taste | design | prompt | anti-slop + taste rules (layout / type / spacing / motion discipline) | | html-mockup-generator | design | hybrid | single-file clickable HTML mockups with phone/browser/macOS frames | | style-minimalist / premium / brutalist | design | prompt | visual-direction switches, auto-discovered by style-pack tag | | brand-asset-protocol | design | hybrid | gather real logo/colors/fonts from official sources → brand-spec, then generate the tokens.json design-token-adherence enforces (deterministic, + optional layered CSS) | | design-system-advisor | design | hybrid | by product archetype (+ optional tone / industry) composes a starter design system from the library's own skills: a style-pack direction, design-taste dials, the anti-patterns to watch, and a neutral starter palette/type to replace with the real brand → JSON or --persist (MASTER.md + pages). A starting point, not dogma. | | skill-lint | meta | deterministic | 11 rules validate the library's own SKILL.md files: frontmatter validity, name=folder, trigger-style descriptions, on-disk link integrity, and a public-cleanliness gate (no leaked email / home paths / blacklisted codenames) → JSON. The library linting itself. |

The detector doubles as an automated gate — see docs/integration.md for pre-commit / CI / render-gate recipes.

The --json output is pinned by a published contract under schemas/ (validate with the bundled zero-dep scripts/lib/validate-output.mjs). Add --remediation to attach each rule's canonical fix to every finding, so an agent can repair mechanically — each detector ships its own reference/remediation.json.

All skills above are zero-dependency except render-verification, which is opt-in: it renders in a real browser, so it needs the optional Playwright peer dependency (npm i -D playwright && npx playwright install chromium; screenshot diff additionally needs pixelmatch + pngjs). Installing the library never pulls a browser; the static detectors stay dependency-free.

Skill format (hybrid)

skills/{category}/{skill-name}/
  SKILL.md      # frontmatter (name/description) + light instruction + when to run a script
  scripts/      # deterministic part (0 tokens): node *.mjs
  reference/    # deep docs, loaded on demand
  assets/       # frames, templates, tokens
  __tests__/    # tests for the scripts

Develop

npm install
npm test          # vitest, run once
npm run test:watch
npm run typecheck # tsc --noEmit (JSDoc + checkJs over .mjs)

Scripts are JavaScript ESM (.mjs). Type-safety comes from JSDoc annotations checked by tsc (no build step — sources run as-is).

Install

Three delivery channels — all install the same zero-dependency skill folders:

# 1. install-CLI -> brand root, served cross-tool by SkillForge
npx @lyupro/llm-skills install
skillforge folders add ~/.lyupro/.skills && skillforge skills reindex

# 2. install-CLI -> native Claude Code (no runtime)
npx @lyupro/llm-skills install --link-claude

# 3. Claude plugin marketplace
/plugin marketplace add lyupro/llm-skills
claude plugin install llm-skills@lyupro

Our skills install under ~/.lyupro/.skills/<name>/ (the leading dot marks Lyu Pro's own content; ~/.lyupro/skills/ is reserved for third-party skills). Full runbook — all channels, options, deploy modes: docs/install-runbook.md.

Stay updated

llm-skills -v               # installed version
llm-skills update --check   # is a newer version published?
llm-skills update           # update to the latest (npm install -g under the hood)

update is fail-loud: on a root-owned global prefix it prints the exact sudo command (never runs it), and an npm min-release-age cooldown is reported with a --min-release-age 0 opt-in rather than bypassed silently. See the runbook for the permission and cooldown details.

License

MIT — see LICENSE. Contact: [email protected]