convez.ai
MCP ServerDesenvolvimento WebIniciante
badhope

AI-SKILL: Índice Curado de Agent Skills

badhope/AI-SKILL

Catálogo curado de skills de agentes IA — 900+ repos em 49 categorias

Ver no GitHub
BomPluginScore · de 0 a 100

Um índice manual de 900+ skills para agentes de IA, organizados em 49 categorias com links diretos para repositórios, cookbooks, MCP servers e estudos de caso. Um único arquivo YAML bilíngue que aponta para projetos upstream verificados, sem reempacotar código alheio.

Instalação

$npx convez-ai badhope/AI-SKILL

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
11
Forks
1
Acompanhando
1
Saúde
71%
Último commit
há 1 semana
Versão
v0.15.0
Issues abertas
0
Contribuidores
3

Casos de uso

  • Descobrir e catalogar skills prontos para agentes Claude
  • Consultar padrões e implementações de tool use, RAG e MCP servers
  • Construir bases de conhecimento sobre o ecossistema de agent frameworks
  • Validar repositórios e estudos de caso antes de integração em produção

Veredito

Ferramenta de referência essencial para quem trabalha com agentes Claude. Não substitui leitura de código, mas poupa horas de busca em GitHub ao concentrar 900+ projetos verificados em um YAML legível. Manutenção manual garante qualidade, não quantidade.

README

A tradução para o português deste README ainda está na fila.

AI-SKILL

A working index of agent skills — repos, cookbooks, MCP servers, production case studies. The stuff you'd actually want when building with LLMs.

Not a library. Not a tutorial site. Just a single YAML file with 900+ curated source_url pointers, sorted into 49 categories. You read it, you click through, you use the upstream project. That is the whole loop.

English · 简体中文

Why this exists

The agent ecosystem has thousands of good repos scattered across GitHub, arXiv, and the corporate research blogs. Searching GitHub directly returns noise. The existing awesome-lists are usually abandoned or vendor-driven.

This repo is one hand-curated YAML file. The value is in the picking: every entry has a source_url that resolves, a one-line summary / summary_zh, and a category from a closed set of 49. You can grep it, vendor it, render it, or clone it as a starting point for your own list.

What's in the box

  • catalog/skills.yaml — the index. 900+ entries across 49 categories, bilingual (en + zh). One entry per upstream repo.
  • catalog/own_skills.yaml — an empty placeholder for skills we author and host in this repo. See skills/ for the format.
  • skills/ — drop-in folder for our own skills (frontmatter + Markdown). Empty by design; fill it when you have something.

We don't host, mirror, or repackage anyone else's work. The catalog points; the upstream delivers.

The 49 categories

The full list lives in catalog/skills.yaml under categories:. A sample so you know the shape:

#CategoryWhat goes here
01Official CookbooksVendor pattern books (Anthropic, OpenAI, Google, …)
03Agent FrameworksLangChain, AutoGen, CrewAI, smolagents, …
04RAG & RetrievalVector stores, retrievers, chunkers, embedding pipelines
10MCP ProtocolModel Context Protocol servers, clients, gateways
22Synthetic DataSelf-instruct, evol-instruct, distillation
35Reasoning ModelsChain-of-thought, o1-style, self-consistency
49Case StudiesProduction post-mortems, industry reports

New entries fit into an existing category. New categories need a reason — open an issue first if you think one is missing.

Entry schema

Each record under the top-level skills: list in catalog/skills.yaml:

- slug: anthropic-cookbook            # unique id, kebab-case
  title: Claude Cookbook              # display name (en)
  title_zh: Claude Cookbook           # display name (zh)
  source: anthropics/anthropic-cookbook         # owner/repo on GitHub
  source_url: https://github.com/anthropics/anthropic-cookbook
  category: official-cookbooks        # one of the 49 slugs above
  subgroup: us-frontier-labs          # optional, for finer grouping
  tags:                               # 2-5 lowercase tags
    - claude
    - anthropic
    - official
    - notebook
  summary: Anthropic's official notebook collection for Claude — tool use, PDF, structured output.
  summary_zh: Anthropic 官方 notebook 合集:工具调用、PDF 理解、结构化输出。
  stars: 22000                        # refreshed daily by sync-github
  license: MIT                        # SPDX id, refreshed daily
  pushed_at: 2026-06-04               # last commit on upstream
  added: 2026-06-04                   # when we added it to the index
  archived: false                     # true if upstream went read-only

source and source_url are the pointers; the rest is metadata. The added date is when the entry landed in the index — don't conflate it with file modification time.

For non-GitHub sources (e.g. arXiv papers, blog postmortems in the case-studies category), the source field can be a domain like arxiv.org and the source_url the actual paper URL. The sync-github script skips those entries automatically.

Using the catalog

The repository is the catalog. There is no docs site, no Pages build, no API. The data is the product.

Programmatic read:

import yaml
with open("catalog/skills.yaml") as f:
    data = yaml.safe_load(f)
entry = next(s for s in data["skills"] if s["slug"] == "langchain")
print(entry["source_url"])

Grep-style read with yq:

# All RAG-related entries, slug + url
yq '.skills[] | select(.category == "rag-retrieval")
    | [.slug, .source_url] | @tsv' catalog/skills.yaml

# All entries from a single org
yq '.skills[] | select(.source | startswith("openai/"))
    | .slug' catalog/skills.yaml

# Categories with their entry counts
yq '.skills | group_by(.category) | .[]
    | {cat: .[0].category, n: length}' catalog/skills.yaml

Extending the catalog

Adding an external skill is one PR, one file edit:

  1. Fork the repo.
  2. Append a new entry under skills: in catalog/skills.yaml using the schema above.
  3. Constraints: slug unique, source_url resolves (2xx, not 404), summary and summary_zh each one line, category from the closed set of 49.
  4. Open a PR. The link-check workflow pings your source_url. After merge, the sync workflow will backfill stars / license / pushed_at within a day.

Publishing your own skill on the in-repo shelf:

  1. Copy skills/_TEMPLATE.md to skills/<your-slug>.md.
  2. Add a row to catalog/own_skills.yaml with at least a path field pointing back to your file.
  3. Open a PR.

No review queue, no formal process. If the entry is right and the link works, it lands.

The two helper scripts

Both are in scripts/. Stdlib-only except for pyyaml and ruamel.yaml. Both work locally and in CI.

scripts/sync-github.py

Refreshes stars, license, pushed_at, and archived for every GitHub-hosted entry via the GitHub API. Runs daily at 06:00 UTC, opens a PR with the diff.

pip install ruamel.yaml
GITHUB_TOKEN=ghp_xxx python scripts/sync-github.py --dry-run   # preview
GITHUB_TOKEN=ghp_xxx python scripts/sync-github.py             # apply

Without a token you get 60 req/h (sequential, ~15 h for 900 entries). With a token you get 5,000/h and 12 concurrent workers (~3 minutes).

scripts/check-links.py

Pings every source_url, classifies the result, and writes the summary to data/health.json. Runs weekly on Mondays at 04:00 UTC. Fails the workflow on any 4xx/5xx/timeout, so dead links are loud.

pip install pyyaml
python scripts/check-links.py            # prints summary
python scripts/check-links.py --fail     # exit 1 on any broken link

If a link breaks (repo moved, went private, the URL is malformed), the script says so and the workflow fails. You fix the URL in the next PR and the next Monday's run goes green again.

When this isn't the right tool

  • You want a deep technical comparison of two frameworks. We don't benchmark; we point you to the upstream.
  • You need real-time star counts. We refresh daily; for live numbers go to GitHub.
  • You're looking for tutorials. We link to the repos that contain them.

License

MIT — see LICENSE. The catalog index is MIT-licensed. Each linked skill keeps its upstream license.

Ver o original no GitHub

Bom para

Relacionados