en zh-CN

Geek Skills

13 curated Claude Code skills that ship finished work — decks, research briefs, PRDs, articles, audits.

End-to-end workflows tested like software, not prompt snippets.

validate skills evals license

Type this —

/deck-studio Turn this quarterly review into a consulting-style deck

— and get a full deck back. Page one of a real run, scored 7.1/10 by an independent blind judge (7.0 = professional design-studio line):

What will you ship today?

I need to… Skill You get
Research a decision properly 🔬 deep-research (v8.1) A cited decision brief — scoped plan → parallel investigation → verified citations
Write or review a PRD 📋 product-manager A PRD your developers can start from, with checkable acceptance bars
Present or pitch 🎞️ deck-studio (v3) A rendered deck: 17-style library, registered layouts, 22-rule visual gate
Publish a Chinese long-form article ✍️ wechat-article-writer A publishable article with a built-in anti-translationese pass

Each is an end-to-end workflow, not a single prompt. All skills ↓

🚀 Install in 30 seconds

git clone https://github.com/staruhub/ClaudeSkills.git && cd ClaudeSkills
python3 scripts/install_skill.py deck-studio      # -> ~/.claude/skills/deck-studio, then run /deck-studio
python3 scripts/install_skill.py --list                  # see installable names
python3 scripts/install_skill.py deep-research           # any skill by short name
python3 scripts/install_skill.py deep-research --project # -> ./.claude/skills/ (project-level)

Manual: the installed directory name is the slash command, so copy and rename:

cp -r skills/Geek-skills-deep-research ~/.claude/skills/deep-research

Copy without renaming → the command becomes /Geek-skills-deep-research. Claude also auto-loads a skill when its description matches; /command is just the explicit way in.

Update / uninstall:

git pull && python3 scripts/install_skill.py deck-studio --force   # update (installed skills are copies)
rm -rf ~/.claude/skills/deck-studio                                # uninstall
  • Installed, but /deck-studio doesn't appear — the slash command comes from the installed directory name. If you copied manually without renaming, your command is /Geek-skills-deck-studio. Re-install with the script, or rename the directory.
  • The skill doesn't trigger automatically — auto-loading matches the skill's description against your request, so phrasing matters. The /command form always works.
  • Do I need to re-install after git pull? — yes: installed skills are copies. Re-run python3 scripts/install_skill.py <name> --force.

📈 Quality you can re-run, not adjectives

Deck quality is scored by independent blind judges on an absolute rubric (10 = design studio, 7 = pro agency) — not by me. The trajectory on the same rubric across four release rounds: 6.0 → 6.6 → 6.6 → 7.1, first past the studio line. In a 3-judge, position-swapped blind eval, the current pipeline beat the previous one 42.3 vs 29.7 — a 42% margin, consistent across swapped positions (method & scores).

Every example directory ships the full generator, rendered pages, and the defects the judges caught: constructivist (7.1) · moshiro (3-judge eval) · yinghuang · polar-night

🧪 Maintained like software

  • Skill Quality Standard v1.0 — every skill passes a D0 gate and carries checkable acceptance criteria, explicit boundaries (when not to use, with hand-offs), and pitfall tables drawn from real failures.
  • Routing evals — 85 cases across 10 skills (skills/*/evals/routing-evals.json) proving each skill triggers when it should and defers when it shouldn't.
  • CI on every pushtwo L1 gates (structure + routing-eval consistency) plus a script compile check. Reproduce locally: python3 scripts/validate.py && python3 scripts/run_routing_evals.py.
  • Know what you installSECURITY.md gives a per-skill capability matrix (reads / writes / network / shells out / credentials / can delete). 9 of the 13 curated skills ship zero code; exactly one can delete files, and it defaults to dry-run.

⚠️ Quality gates are a self-audit by Claude, not third-party certification — the commands above let you re-run them yourself. Full refactor record: CHANGELOG.md.

📚 All Skills

Flagship — the four end-to-end workflows above: deck-studio · deep-research · product-manager · wechat-article-writer

Skill Description
pair-programming Pair-programming partner: delivers code with a structured self-review, focused on AI-specific defects
security-audit Comprehensive code security audit
solution-architect System design, tech selection, and architecture review
threejs-performance Three.js performance optimization
mineru-pdf-parser PDF to Markdown or JSON for LLM workflows (requires a local MinerU install)
ai-sales-champion AI sales/consulting dialogue helper — turn tech into business language
keqian-method Keqian's AI-Native product dev methodology: single-agent, SDD, quality gates
xuefeng-method Xuefeng's AI-Native methodology for open-behavior, model-driven products
c-drive-cleaner Windows C drive cleanup and disk space management (dry-run by default)

Lab — experimental and personal skills (exam prep, weather reports, image/podcast generation, A-share analysis) live in lab/. They are not part of the curated set, are excluded from the quality gates above, and may graduate into skills/ or move out of this repo.

Skill Notes
llm-wiki Codebase wiki builder, based on Karpathy's LLM Wiki pattern; kept in its original upstream layout at repo root

🤝 Community

Found a bug, or built something with a skill? Open an issue. Want to contribute a skill or a fix? Start with CONTRIBUTING.md — new skills incubate in lab/ and graduate into the curated set by passing the quality gates. If a skill saved you an afternoon, a ⭐ helps other people find it.

License

MIT