Qiaomu Meta Skill

Qiaomu Meta Skill

Build reusable Qiaomu skill packages, not long prompts.

Router Rules

• Route by frontmatter description first.
• Once selected, qiaomu-meta-skill is the single authoring authority. Do not also invoke a generic skill-creator unless the user explicitly requests comparison or this skill is unavailable.
• Built-in prior-art discovery belongs to this skill. Do not install, load, or delegate to a separate discovery skill.
• Built-in GitHub publishing belongs to this skill. Do not require or invoke a separate publisher skill after this package is selected.
• Keep the package root SKILL.md to routing and the minimal workflow. Put judgment in references/, deterministic behavior in scripts/, regression cases in evals/, and evidence in reports/.
• A package has one discoverable root SKILL.md; embedded examples and fixtures use SKILL.example.md or SKILL.fixture.md.
• Do not turn one-off summaries, translations, explanations, or brainstorming into skills.
• Match the user's action: create/refactor/package requests may edit; audit/evaluate/diagnose-only requests remain read-only; publish only when explicitly requested.
• Default to concise Chinese-first qiaomu- names with no more than three preferred hyphen parts.
• Add Copyright (c) 向阳乔木, X https://x.com/vista8, and GitHub https://github.com/joeseesun/ unless another owner is requested.

Modes

• Scaffold: exploratory or personal; minimum useful files.
• Production: team reuse; README, interface, trigger eval, output contract, and install evidence.
• Library: shared infrastructure; Production plus Skill IR, portability, trust, and review cadence.
• Governed: public or high-trust; Library plus permission, rollback, secret, release, and claim gates.

Choose proportionally with Operating Modes, Gate Selection, and QA Ladder.

Built-In Prior-Art Discovery

Before a new skill or substantial redesign:

1. Derive 2–4 intent-shaped queries covering outcome, domain action, quality mechanism, and an adjacent synonym when useful.
2. Prefer the unified runner:

python3 scripts/research_prior_art.py "<query 1>" "<query 2>" --strict --summary --output reports/prior-art-candidates.json

Its underlying catalog calls remain:

npx --yes skills find "<query>"
python3 scripts/search_skillsmp.py "<query>" --limit 20 --sort stars

3. Keep metrics separate: skills.sh installs measure adoption; SkillsMP stars belong to the source repository; neither is a user rating or quality score.
4. Deduplicate by canonical GitHub repository and skill path. Collapse translations, mirrors, and obvious forks without adding metrics together.
5. Shortlist genuinely relevant popularity, trust, and complementary anchors. Inspect source SKILL.md, maintenance, license, permissions, security signals, and available rating evidence; never execute untrusted candidate code just to study it.
6. Synthesize keep / adapt / reject / invent. Map each adopted mechanism to the new package instead of collaging prose.
7. Preserve dated sources, metrics, failures, deduplication, lessons, rejections, and missing evidence in reports/prior-art-research.md for Production+ or materially researched work.

If a catalog fails, continue with the other sources, record missing evidence, and lower the claim. Full method: Prior-Art Research.

Generalization Gate

Before promoting one failure into a core rule:

1. restate it as a domain-neutral behavior
2. classify it as core mechanism, optional adapter, or eval-only fixture
3. promote only safety/factual/permission invariants or behavior repeated across unrelated domains
4. keep one-off details in fixtures or specialist references
5. rerun the original and unrelated boundary cases

Prefer intent fidelity, source fidelity, and decision rules over an expanding topic encyclopedia.

Qiaomu Skill OS

1. Intent: recurring job, users, inputs, output, exclusions, standards, references.
2. Skill IR: platform-neutral meaning and evidence boundary.
3. Package: lean root instructions, interface, README, and earned resources.
4. Eval: trigger boundaries first; output/runtime/human eval when risk justifies it.
5. Review: package, context, trust, install, README, and public claims.
6. Operate: explicit feedback, failures, drift, and next-iteration proposals without raw private content.

Compact Workflow

1. Decide whether the request deserves a reusable skill; otherwise answer directly and create no package.
2. Capture job, finished output, target users, inputs, exclusions, permissions, standards, existing assets, platforms, and publication intent.
3. Pass prior-art discovery or record why it is not applicable or missing evidence.
4. Pass the generalization gate for sample-driven core changes.
5. Choose the lightest valid mode.
6. Write the description early; run evals/trigger_cases.json before expanding structure.
7. Create only earned resources. Never create ceremonial directories or duplicate README/SKILL prose.
8. Export reports/skill-ir.json for Production+, public, or cross-platform packages.
9. Add output evals when correctness, safety, persuasion, or repeatability cannot be shown by trigger tests alone.
10. Keep mutations within the requested action boundary and preserve rollback for risky changes.
11. Validate package, unit tests, trigger behavior, context budget, secret/trust boundaries, and evidence claims.
12. Produce the creation handoff and clearly label missing evidence.
13. When publication is requested, read Self-Contained Skill Publishing, then use the bundled publisher for feature branch → validation → PR → merge → release/install verification; never push directly to the default branch.

Core commands:

python3 scripts/validate_skill.py .
python3 scripts/export_skill_ir.py . --output reports/skill-ir.json
python3 scripts/trigger_eval.py . --cases evals/trigger_cases.json --output reports/trigger-eval.json
python3 scripts/release_check.py . --phase local --run-tests
python3 scripts/publish_skill.py /path/to/skill --dry-run