Yao Meta Skill
Turns repeated workflows, prompts, and SOPs into installable, evaluated, governed cross-platform agent skill packages.
Terms like "governance", "trust report", and "output_quality_scorecard" read more like self-description of a framework than verifiable mechanisms; the referenced files weren't retrieved to confirm details.
The main doc is extremely dense and jargon-heavy ("Skill IR", "trust report", "Governed Package Boundary"), pushing concrete rules into references/ files that weren't fetched in this review — hard to tell how to actually use it from this file alone.
- The main document is very dense and defers concrete rules to references/ files not retrieved in this review — actual usability is unverified
What does this skill do, and when should you use it?
yao-meta-skill is a meta-skill for creating, improving, evaluating, and packaging reusable agent skills. It uses SKILL.md as its entrypoint, backed by a unified CLI (scripts/yao.py) and method doctrine in references/. The project evolved from a 1.0 line (turning workflows into installable packages) into Skill OS 2.0: model a skill once, compile to multiple targets, gather evaluation evidence, gate releases, and track operations. The authors' self-scored weighted review gives it 91.5/100, though stronger public claims remain explicitly evidence-gated.
Reads raw workflow notes, prompt sets, transcripts, or runbooks, clarifies the job through a short intent dialogue, and generates a skill package (SKILL.md, agents/interface.yaml, references, scripts, evals, reports). It provides: train/dev/holdout trigger evaluation (currently 66 prompts across 21 families at 1.0 precision/recall); a Skill IR intermediate representation with target compilers for OpenAI, Claude, generic Agent Skills, and VS Code; an Output Eval Lab with assertion grading, blind review packs, and adjudication reports; a Review Studio HTML release-gate page; and package verification, install simulation, runtime permission probes, trust reports, and evidence-consistency checks. It routes by frontmatter description; self-updates require python3 scripts/yao.py self-update --self --yes.
- An internal tooling team wants to convert a repeatedly executed workflow note into an installable package with explicit boundaries and evals.
- A skill author wants to promote a personal skill into a team asset with interface contracts, trust checks, and review evidence before others depend on it.
- A maintainer is preparing a beta release and needs package verification, install simulation, and compatibility checks, keeping beta readiness separate from stronger claims.
- After release, an operator uses adoption drift, metadata-only telemetry, and weekly SkillOps reports to decide between documentation, an eval, a patch, or a governance update.
- A builder comparing meta-skill approaches wants evidence, portability, and release gates rather than conversational-only creation.
What are this skill's strengths and limitations?
- Rigorous evaluation: trigger quality is checked with family regressions, blind/adversarial holdout, and an independent rubric judge rather than intuition.
- Strong portability design: Skill IR stays platform-neutral; openai/claude/generic/vscode targets compile cleanly with a 100/100 portability score.
- Governance is runnable, not prose: ownership, review cadence, maturity scores, release locks, and drift detection all have executable checks.
- Evidence is published and reproducible: regression corpora, failure library, drift history, and promotion decisions ship as artifacts.
- High complexity: dozens of subcommands and report artifacts create a much steeper learning curve than lean skill-writing methods.
- Stronger claims remain evidence-pending: no provider-backed independent model execution evidence or real-client telemetry yet.
- The human blind A/B comparison is a single reviewer over 5 scenarios with empty per-case rationale fields.
- The weighted benchmark (91.5 vs Anthropic 67.5 / OpenAI 50.5) is a self-built, self-scored comparison and is inherently subjective.
- All current target adapters are metadata fallbacks (4/4) with no native permission enforcement.
How do you install this skill?
Install globally for Codex: npx -y skills add yaojingang/yao-meta-skill -a codex -g -y; for all supported agents replace -a codex with -a '*'. Restart the client after installation. Local development requires Python 3.11+; run tests with make ci-test. The source does not document other installation routes.
How do you use this skill?
After install and restart, trigger it with prompts like "create a skill from this workflow", "improve this existing skill", or "add evals to this skill". It runs a short intent dialogue, then a silent benchmark scan, and generates the package. Alternatively use the unified CLI: python3 scripts/yao.py quickstart --output-dir .; the bilingual interpretation report lands in reports/skill-interpretation.html, the scorecard in reports/skill-overview.html, and the release gate in reports/review-studio.html.