| name | yao-meta-skill |
| description | Create, improve, or evaluate an existing skill from workflows, prompts, SOPs, scripts. Use for migration/release/package, routing, evals/tests, install/trust checks, 优化已有 skill, 补 trigger 评测. Exclude summary/translation/docs, brainstorming, one-off prompts, copy-only edits, outline-only requests. |
| metadata | {"author":"Yao Team"} |
Yao Meta Skill
Router Rules
- Route by frontmatter
description using the lightest mode.
- Keep
SKILL.md lean: guidance in references/, logic in scripts/, evidence in reports/.
Modes
Scaffold: exploratory; Production: team; Library: shared; Governed: release-critical.- Rules: Method, Modes, Boundaries.
Compact Workflow
- Require
repeated use + reusable output contract; skip one-offs.
- Lock target path + identity before writes; self-edits require
--self.
- Capture job, output, exclusions, constraints, standards, mode.
- Scan
3-5 useful external/user/local references.
- Write
description early; route edits need trigger_eval.py; release gates follow risk.
- Add deeper profiles only when earned.
Playbooks: Target Safety, Intent, Skill IR, Output Eval, Review Studio.
Updates
Activation: scripts/yao.py check-update --notice --self; show notice_text if requested; continue. “更新”: scripts/yao.py self-update --self --yes.
Skill OS 2.0 Gates
For production/library/governed releases, run Skill IR, compiler, trigger/output eval, Atlas, conformance, trust, package/install, upgrade, drift, waiver, Review Studio.
Governed Package Boundary
Governed/file-backed packages: label input_files a file-backed fixture; include owner, review cadence, output contract, rollback boundary; require trust report and reports/output_quality_scorecard.md; mark unavailable telemetry/approvals/metrics/benchmarks as missing evidence; never fabricate evidence.
Preserve labels literally when they apply: file-backed fixture, input_files, output contract, rollback boundary, trust report, reports/output_quality_scorecard.md, missing evidence.
First-Turn Style
- Infer non-core gaps; ask one core-fork question per round, max two, then use recorded preferred inference.
- In Chinese, sound soft and companion-like; use Intent Dialogue.
Output Contract
Builds: SKILL.md, aligned interface, justified assets, boundary/gate summary. Audits: findings + fixes; edit only when asked. No-skill: no files.
Reference Map
Primary: Method, Artifact Design, Systems, Governance, SkillOps.
