---

Quick Start

npx skills add LearnPrompt/luban-skill -g

Claude Code users can also install via the plugin marketplace (auto-updates):

/plugin marketplace add LearnPrompt/luban-skill
/plugin install luban

Once installed, tell your Agent:

Let Luban take a look at my skill: [your Skill directory / GitHub repo link / SKILL.md content]

Luban will first complete material inspection, trade survey, positioning, and measurement, then give you three polishing directions with a recommendation — it will not touch a single line until you pick a direction.

What Problem It Solves

You wrote a Skill. It works fine for you. Then what?

• You publish it on GitHub and nobody installs it — people can't tell what it does;
• The README reads like an engineering spec: no first-screen hook, no screenshot-worthy output;
• You say it "works well," but you can't produce a single reproducible piece of evidence;
• You want to improve it, but don't know where to start — rewrite the trigger phrases? Redo the workflow? Add a showcase?

The usual approach is "polish this up for me." Luban's approach is to take it into the workshop as a piece of craftsmanship: first challenge whether it's worth carving at all, then see how its peers earn their place, measure its weak spots with three rulers, plane it stroke by stroke — every stroke must pass a validation gate — and finally ship a release with house rules set in place.

What It Delivers

• A 13-section Polishing Report: material-inspection verdict, peer comparison table (all with URLs), ecological-niche assessment, scorecard, three polishing directions, and drop-in rewrite snippets
• A screenshot-ready "Graduation Certificate": before/after scores, a one-line new positioning, the signature strength, and next steps
• Verification assets deposited into your repo: one-off comparison scripts hardened into tools, and judgment criteria written down as explicit project rules

The Five Moves

Move	What it does	One blunt line
验料 Yanliao / Inspect Materials	Challenge whether the Skill's premise holds at all	Rotten wood cannot be carved — if it's not worth it, say so
访行 Fangxing / Survey the Trade	Search the web for peers and map the ecological niche	No good tool was ever built behind closed doors
过尺 Guochi / Measure Up	Score it with three rulers: structure, live testing, liveness	Green CI can lie — reconcile against real artifacts
慢刨 Manpao / Plane Slowly	Freeze the baseline; keep a change only if it passes the validation gate	If it doesn't measure up, undo the stroke — never plane just to look busy
回炉 Huilu / Back to the Forge	Leave a post-release watchlist; next round starts from feedback	Delivery is not the finish line

How to Trigger

• "Let Luban look at this skill" / "Polish it at the master's gate"
• "Upgrade my skill" / "Polish my skill"
• "Skill checkup" / "Skill audit"
• "Why does nobody install my skill"
• "How do I publish this skill to GitHub/ClawHub"
• "Benchmark this against similar skills"

Track Record

Luban's first job: polishing ai-news-radar (~1k stars) from v0.6 to v0.7.0, completed within a single conversation, with all 4 PRs merged:

• The liveness check caught a data pipeline that had been silently stale for 8 days under all-green Actions (a git add allowlist omission)
• The scoring fix was validated by replaying 83,725 historical records: 327 false-AI entries purged, zero collateral damage
• Single-source share in the featured section: 15/20 → 4/20; first-screen render: 806 → 523 cards, page height -30%
• The verification method was deposited as a repo tool, backtest_scoring.py, along with a new project rule: "any scoring change must ship with a ≥14-day replay"

Full record (including incidents and lessons): skills/luban/examples/ai-news-radar-case.md — every number links to a clickable PR.

How It Differs from the Alternatives

	Just asking an Agent to "make it look nicer"	Luban
Starting point	Rewrites the copy immediately	First challenges the premise: does this Skill deserve to exist?
Basis	The model's taste	Peer benchmarking (with URLs) + three-ruler scoring (with evidence)
Method	Changes everything at once, impossible to attribute	Freezes the baseline; one commit per facet; kept only if it passes the validation gate
Verification	"Looks better"	Real-data replay with before/after flip numbers
Ending	Done when the edits are done	Verification tools deposited, rules established, forge-return watchlist left behind

Safety Boundaries

• Mandatory stop points: proposing a repositioning, merging to the default branch, tagging a release, any deployment visible to real users — each waits for your explicit authorization; your question ("all good, right?") does not count as authorization.
• Never writes API keys, tokens, cookies, or private paths into any public artifact.
• Changes are always presented as auditable commits — no brute-force rollbacks like git reset --hard.

File Structure

luban-skill/
├── skills/luban/
│   ├── SKILL.md                    # The workflow itself: five moves, nine-step process, house rules and acceptance checklist
│   └── examples/
│       └── ai-news-radar-case.md  # Case study: real repo, real numbers, fully verifiable
├── assets/                         # demo GIF and reproducible recording script (vhs tape)
├── .claude-plugin/                 # Claude Code plugin marketplace manifest
├── README.md
└── LICENSE

Verification & Testing

After installing, run this acceptance check:

Let Luban look at this skill: https://github.com/anthropics/skills

A passing run: it first outputs the material-inspection challenge and a peer comparison with URLs, offers three directions, and stops to wait for your choice — instead of jumping straight into rewriting.

Methodology Credits

Luban's five moves weren't invented from thin air; they stand on the shoulders of:

• KKKKhazix/khazix-skills · hv-analysis — Horizontal-vertical analysis: trace the lineage vertically, compare peers horizontally, and let the intersection form the judgment (the skeleton of Fangxing and positioning)
• alchaincyf/darwin-skill — Evaluate → improve → test → keep or roll back; independent-judge perspective; the ratchet mechanism (the soul of Guochi and Manpao)
• microsoft/SkillOpt — Frozen baselines, bounded candidate edits, validation-gated acceptance (the origin of the validation gate)
• And one real end-to-end engagement, ai-news-radar v0.7.0 — liveness checks, verification-asset deposits, workbench discipline, and the return to the forge were all learned there

License

MIT — use it, modify it, let it polish your work.

---

---