Create, refactor, and package reusable agent skills

Yao Meta Skill creates, refactors, evaluates, and packages agent skills from workflows, prompts, transcripts, docs, or notes - covering skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution. It routes by frontmatter description and keeps SKILL.md lean, pushing guidance into references/, logic into scripts/, and evidence into reports/, always favoring the lightest reliable process.

Use it when the request matches creating, refactoring, evaluating, or packaging a reusable agent skill. For a one-off task with no reusable process, the compact workflow's first step is explicit: do not create a skill - check for a near-neighbor skill first, and only proceed if there's genuinely repeated use plus a reusable output contract.

It defines four operating modes of increasing rigor - Scaffold (exploratory/personal), Production (team reuse), Library (shared infrastructure), and Governed (high-trust, policy-sensitive, or release-critical) - governed by dedicated Method, Operating Modes, and Resource Boundaries reference docs. Its compact workflow captures the job, output, exclusions, constraints, and standards; scans external benchmark, user source, and local-fit references in that order, surfacing only genuine uncertainty or conflict; writes the description early and tests route quality before adding folders or gates; and layers in output-risk, artifact-design, prompt-quality, system-model, or next-directions guidance only when useful. Unless told otherwise, the output contract is SKILL.md, an aligned agents/interface.yaml, justified supporting assets, and a short summary of boundary, exclusions, gates, and next steps.

# agents/interface.yaml is produced alongside SKILL.md

For Production, Library, Governed, or team-distributed work, it runs a Skill OS 2.0 gate chain before release - Skill IR, target compiler, trigger and output eval, Skill Atlas, conformance, trust, registry/package/install, upgrade, drift, waiver, and Review Studio. For file-backed, release-critical, or governed packages it requires an owner, review cadence, input_files framed as file-backed fixture evidence, an output contract, and a rollback boundary, plus a trust report and reports/output_quality_scorecard.md - explicitly marking any unavailable telemetry, approvals, metrics, or benchmarks as missing evidence rather than fabricating it. Its first-turn style starts from the user's actual work or outcome before structure, asks only 2-3 key questions unless enough detail already exists, and in Chinese adopts a soft, companion-like tone via a dedicated Intent Dialogue reference.

Teams and individuals who need to turn ad hoc workflows, prompts, transcripts, or notes into properly scoped, evaluated agent skills - scaling the process from a quick personal scaffold up to a governed, audit-ready package with evidence and rollback boundaries when the stakes require it.