Turn Repeated Workflows Into Governed, Team-Ready Skills

A team's operating knowledge rarely lives in documents — it lives in chat history, personal prompts, and habits nobody wrote down. yao-meta-skill (YAO stands for Yielding AI Outcomes), a 3k-star MIT project, is a meta-skill: a skill whose job is creating, evaluating, packaging, and governing your other skills.

Why This Skill Matters

The 1.0 line turned repeated workflows into installable skill packages. The 2.0 line, which the README calls a Skill OS, wraps the generated SKILL.md in a governed lifecycle:

• Skill IR — a platform-neutral intermediate representation for intent, triggers, inputs, outputs, boundaries, references, and expected artifacts.
• Target compilers and adapters — the same source compiles to OpenAI, Claude, generic agent-skill, Agent Skills-compatible, and VS Code-oriented surfaces.
• Output Eval Lab — trigger checks, output assertions, execution evidence, and blind-review packs.
• Review Studio 2.0 — a single HTML gate page combining intent, trigger, output eval, context cost, trust, and release evidence.
• Evidence and release governance — package verification, install simulation, and a claim guard that blocks premature "world-class" claims while evidence tasks are pending.
• SkillOps loop — adoption drift and weekly curator reports that feed the next iteration.

Installation

Install globally for Codex:

npx -y skills add yaojingang/yao-meta-skill -a codex -g -y

To install for every supported agent, replace -a codex with -a '*'. Restart the client afterward, then ask naturally — the README lists trigger phrases such as "create a skill from this workflow", "improve this existing skill", "evaluate this skill", or "add evals to this skill". Developing the factory itself wants Python 3.11 or newer.

Real Workflow: A Workflow Note Becomes a Governed Package

The README's quick-start flow, run through the unified yao.py CLI:

1. Describe the workflow you want to capture. The intent dialogue asks one personalized question only for core forks — task, output, or direction — stops after two rounds, and records structured assumptions.
2. quickstart runs a silent benchmark scan and reference synthesis; it surfaces questions only when intent is unclear or a real design conflict exists.
3. Generate the package in one of the archetype-aware modes: scaffold, production, library, or governed.
4. Review three HTML reports: reports/skill-interpretation.html (defaults to Simplified Chinese, with an English switch in the top right), reports/skill-overview.html for the audit scorecard, and reports/review-studio.html for release blockers, permission approvals, and evidence paths.
5. Validate and review:

python3 scripts/yao.py init my-skill --description "Describe what the skill does."
python3 scripts/yao.py validate my-skill
python3 scripts/yao.py review-viewer my-skill

When the package is ready for other clients, the README's export path compiles the Skill IR into target-specific contracts and packages them — compile-skill and package take --target / --platform flags for openai, claude, generic, and vscode.

Tips

• Trigger quality is the core discipline: the README's validation notes describe train/dev/holdout suites, blind and adversarial holdouts, route-confusion tracking, and a judge-backed blind eval before a description is promotable.
• Keep the entrypoint lean — details belong in references/ and scripts/. Context budgets are explicit and tracked; the README reports the root skill at 944/1000.
• Treat the README's weighted benchmark as self-assessment: it scores Yao 91.5/100 against Anthropic's and OpenAI's skill creators, with stars deliberately excluded — it is the project scoring itself, not third-party evidence.
• The 5/5 human blind A/B result over OpenAI's skill-creator is a self-published snapshot from a single reviewer on 2026-06-29, and the README itself labels that boundary.
• Runtime update checks are cached and read-only: python3 scripts/yao.py check-update --notice --self checks the GitHub VERSION at most once every 24 hours and leaves the active skill task unchanged.

When Not to Use This

A one-off prompt does not need a package — the repo ships a Non-Skill Decision Tree doctrine for exactly that call. The README also positions conversational creation (Anthropic/OpenAI style) as the right tool when human-guided iteration matters more than repository-level governance, and suggests a hybrid: draft conversationally, then harden with yao when the package needs evidence, portability, and release gates. And the governance machinery has a cost — for a personal snippet nobody else will run, the README's own design rule is to add extra structure only when it pays for itself.

See the leaderboard for more skills.