Library / SDK
yaojingang/yao-meta-skill avatar
yaojingang/yao-meta-skill

yao-meta-skill ships 71 named checks, no release to install

YAO = Yielding AI Outcomes. A rigorous engineering, evaluation, governance, and portability system for reusable agent skills.

2,732 stars255 forksPythonMIT

At a glance

What is it?
yao-meta-skill turns repeated workflows into installable agent skill packages, then wraps them in what it calls a Skill OS: an intermediate representation, target compilers, an eval lab, review gates and telemetry. What it does not have is a distribution story. There is no GitHub release, no package index, a VERSION file, and a Makefile whose install paths point into the agent skill directories inside your home directory.
Who is it for?
yao-meta-skill fits a team that already writes skills by hand and now needs evidence, portability and a release gate around them, and the project's own framing is the fairest reading of it: beta ready for external testing, with its larger public claims still waiting on evidence it names rather than assumes. Before adopting it, check three things.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 6 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • yaojingang/yao-meta-skill2,732 ★ · +123
Star history
Repository2026-09-102026-11-15
yaojingang/yao-meta-skill2.6k2.7k

123 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The repository is shaped like the package it produces

The root of this project is arranged as a skill package rather than as a library: SKILL.md sits at the top next to manifest.json, VERSION, a Makefile and AGENTS.md, and the directories around them are the ones a package is built from, including agents/, references/, registry/, schemas/, templates/, evals/, runtime/, security/, skill-ir/, skill_atlas/, evidence/, failures/, reports/, examples/ and tests/.

Distribution is the part that is not settled. There is no GitHub release to download and no index to install from, and no packaging line appears in the instructions: no pip command, no index name. A VERSION file stands in for a tag, requirements-ci.txt stands in for declared dependencies, and the Makefile carries the interpreter constraint through `PYTHON ?= python3` and `REQUIRED_PYTHON_VERSION ?= 3.11`, both overridable on the command line. Every operator command starts with `python3 scripts/yao.py`, and the first one takes `--expected-source .`, which reads as an assumption that you are standing inside a checkout.

Two install directories under the home directory, one of them disabled

The Makefile defines two install paths that differ by a single directory level:

makefile
LOCAL_SKILL_INSTALL_DIR ?= $(HOME)/.agents/skills.disabled/yao-meta-skill
ACTIVE_SKILL_INSTALL_DIR ?= $(HOME)/.agents/skills/yao-meta-skill

One lands in an active location, the other in a `.disabled` mirror beside it. The operator command `install-status` is the other half of that arrangement: it reports whether the active skill is coming from `.codex/skills`, from `.agents/skills` or from the disabled mirror, and it flags duplicate active installs. So the tool has two host conventions to reconcile rather than one, and treats two live copies as something worth reporting rather than something to resolve.

In the Makefile's own check list, the only target with install in its name is install-simulation-check, which belongs to the release verification group. Putting files on disk is left to the host's skill loader and to the person holding the checkout.

Seventy one named checks, and a name that stops halfway

The `.PHONY` line names 71 targets, nine of them carrying a world-class prefix and four more named for telemetry. They read as a verification matrix rather than a task list: route-scorecard, route-confusion-check, description-optimization, judge-blind-eval, promotion-check, python-compat-check, target-safety-check, html-rendering-check, runtime-conformance-check, runtime-permission-check, trust-check, registry-audit-check, package-verify-check, review-studio-check, benchmark-reproducibility-check, evidence-consistency-check, adoption-drift-check, telemetry-import-check, telemetry-emit-check, review-waivers-check, intent-confidence-check, output-risk-profile-check.

The inventory does not close. The list stops partway through a name, at system-model-che, and what the top-level build script shows ends there, so anyone counting the project's checks is working from a partial list and the last name they read is half a word. The grouping is also uneven: the release gate targets, the review page targets, the telemetry targets and the evidence targets all sit on one line with nothing separating a fast local check from a full evidence audit.

The project names its own limits and gates its loudest claim

Two passages do the opposite of marketing. One use case is titled Compare with other meta-skill approaches, and it says to keep Anthropic and OpenAI style conversational creation and lean instruction writing where they fit, then reach for yao-meta-skill when the package needs evidence, portability, release gates and repeatable maintenance. That is the project's own drawing of the line between a conversation that writes a prompt and a system that ships a package other people install.

The second is a posture statement: the repository is ready for beta and external testing, while stronger public world-class claims remain evidence-gated. Four items are named as tracked evidence tasks rather than completed work, namely provider-backed production evidence, human blind-review evidence, native permission execution and real-client telemetry. Nine Makefile targets carry the same world-class prefix and world-class-claim-guard-check sits among them, which is the mechanism meant to keep any future claim tied back to that ledger.

Five translations, one sync check, and a badge row of empty links

The documentation exists in five languages in the tree: README.md beside docs/README.zh-CN.md, docs/README.ja-JP.md, docs/README.fr-FR.md and docs/README.ru-RU.md. The operator command `localized-doc-sync-check` verifies one relationship between two of them, that the Chinese README carries the public homepage sections added to the English one. None of the visible commands covers Japanese, French or Russian, so three of the four translations have no check behind them.

The row of links at the top of the file is six entries with empty link text, and only one of them is a badge in the usual sense. The first points at the test.yml workflow, and the other five point at LICENSE, README.md and the four translated READMEs. So the leading element of a badge row is a continuous integration link, and four of the six entries exist to make the translations clickable rather than to report a build state.

Three helper commands, one of which refuses to touch the pull request

The operator section offers three commands and calls them read-only:

bash
python3 scripts/yao.py install-status --expected-source .
python3 scripts/yao.py localized-doc-sync-check
python3 scripts/yao.py pr-review-report 4 --repo yaojingang/yao-meta-skill

The third takes a pull request number and a repository, reads GitHub pull request metadata, changed files, status checks and suggested local commands, and stops there, without merging or mutating the pull request. That limit is the part worth knowing, because a tool that reads a review request usually carries the credentials to answer it too.

The other two are narrower than their names imply. The sync check compares one language against another rather than sweeping all five, and install-status reports where the loaded copy of the skill sits instead of installing anything. The commands are diagnostics for a maintainer who already has a checkout, not a first step for a new user.

The capability list ends mid-word, after three complete bullets

The last section in the file promises to turn rough workflows, transcripts, prompts, notes and runbooks into packages with a trigger surface, a lean SKILL.md and optional references, scripts and evals. Its fourth bullet describes an inference-first intent dialogue: it asks one personalized question, only for core task, output or direction forks, and it stops after two rounds. Then the line ends on the word struc, in the middle of the word, and the file stops there.

That makes the interaction contract the least complete part of the page, and it is the part that decides how a skill asks for its inputs. The navigation line at the top promises Quick Start, Benchmark, Method Doctrine, Evals and a Failure Library; what follows it in this file is the 2.0 upgrade, the 1.0 to 2.0 table, the use cases, the three operator commands and these four bullets. The comparison table is the one place where the two generations sit side by side, in seven rows running from product role down to the operating loop.

Editorial conclusion

yao-meta-skill fits a team that already writes skills by hand and now needs evidence, portability and a release gate around them, and the project's own framing is the fairest reading of it: beta ready for external testing, with its larger public claims still waiting on evidence it names rather than assumes. Before adopting it, check three things. Decide where the skill is meant to live, because the Makefile writes to ~/.agents/skills and the install-status command also expects a .codex/skills layout and flags two active copies. Count the checks you will actually run, since the list of 71 named targets is a verification matrix, not a quick start. And read the review studio and evidence sections in full before treating any benchmark number as a result, because the file itself says the strongest claims are gated.

Frequently asked questions

Can you give me an example of a meta skill?

This repository ships five of them under examples/, indexed by examples/README.md: simple-note-cleanup, team-frontend-review, evolution-frontend-review, complex-release-orchestrator and governed-incident-command. Each is a directory in the tree rather than a link to someone else's project.

What is yao-meta-skill used for?

It creates, evaluates, packages and governs reusable agent skills. The 1.0 line turned repeated workflows into installable, readable, cross-platform skill packages; the 2.0 line adds what the file calls a Skill OS around that, with a Skill IR, target compilers, an output eval lab, a review gate page and a SkillOps loop.

Can I install yao-meta-skill from a release?

There is no GitHub release on this repository, and no index or pip line in the instructions. A VERSION file at the root holds the version, requirements-ci.txt carries dependencies, and the Makefile pins the interpreter through `REQUIRED_PYTHON_VERSION ?= 3.11` while offering 71 named check targets to run.

Where does yao-meta-skill put files on my machine?

The Makefile defines `ACTIVE_SKILL_INSTALL_DIR ?= $(HOME)/.agents/skills/yao-meta-skill` and a disabled mirror at `$(HOME)/.agents/skills.disabled/yao-meta-skill`. The install-status command also reports on `.codex/skills` and `.agents/skills` and flags duplicate active installs, so two host layouts are expected.

Which agent platforms can a yao-meta-skill package target?

The 2.0 target compilers and adapters generate surfaces for OpenAI, Claude, generic agent skills, Agent Skills compatible packages and VS Code-oriented workflows. The intermediate representation is described as platform neutral, covering intent, triggers, inputs, outputs, boundaries, references and expected artifacts.

Official sources

  1. Issues
  2. License: MIT
  3. Project website
  4. README
  5. yaojingang/yao-meta-skill on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/yaojingang-yao-meta-skill.svg)](https://hysenlabs.com/projects/yaojingang-yao-meta-skill)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/yaojingang-yao-meta-skill.svg)](https://hysenlabs.com/projects/yaojingang-yao-meta-skill)

Was this useful?