convez.ai
MCP ServerDevOps & InfraestruturaAvançado
JeremyKuhne

Agent Skills - Repositório de Skills Compartilhadas

JeremyKuhne/agent-skills

Skills portáveis compartilhadas para agentes em .NET com versionamento

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server em PowerShell que gerencia skills reutilizáveis para agentes Claude/Copilot em repositórios .NET. Funciona como um commons centralizado que permite versionar, distribuir e consumir skills genéricas com overlays específicas por repositório.

Instalação

$npx convez-ai JeremyKuhne/agent-skills

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
1
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 1 semana
Versão
v0.17.0
Issues abertas
3
Contribuidores
2

Casos de uso

  • Distribuir skills reutilizáveis entre múltiplos repositórios .NET
  • Gerenciar versões pinadas de skills com validação de proveniência
  • Criar repositórios de skills personalizados com workflows de revisão

Veredito

MCP server bem arquitetado para cenários enterprise de reutilização de skills em portfólios .NET, com modelo bidireccional robusto (core genérico + overlay específico). Requer familiaridade com gh CLI e fluxos de versionamento, adequado para times que gerenciam múltiplos repositórios.

README

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

agent-skills

Shared agent skills and MCP configuration for Jeremy Kuhne's .NET repositories. This is the commons: the single upstream source from which each repository vendors pinned, provenance-stamped copies of the skills it needs.

It is both a gh skill source and a GitHub Copilot CLI plugin marketplace.

Status: pre-1.0 and in active fleet use. The repository currently ships 24 shared skill cores and two reviewer agents. Deterministic source, isolated-install, plugin, scaffold, and synthetic-consumer gates are in place. Model-evaluation vertical slices cover pull-request publication, technical writing, user voice, and skill-repository scaffolding. An opt-in domain suite covers .NET pipes; portfolio-wide routing and outcome coverage remain future work.

The model

Skills are authored once as a portable core and shared from here. Each consuming repository (touki, madowaku, thirtytwo, ...) holds a vendored copy plus a thin repo-specific overlay:

  • The core is generic: no repo-specific paths, project names, cross-references, or source-example links. It must pass a consumer's link check unchanged.
  • The overlay lives in the consuming repo and carries everything specific to it - paths, cross-references to that repo's other skills, example links.

The commons is bidirectional: a generic improvement made in any repo flows back here (when plausible - upstreaming is never automatic), while a repo-specific tweak stays in that repo's overlay. The lifecycle is driven by the manage-skills skill (find / build / review / update / retire).

Creating your own skill repository

If you want to create a skill repository, clone or open this repository in GitHub Copilot and ask that question. The repository-local create-skill-repo workflow will:

  1. explain how original skills are published and how other repositories use version-pinned copies;
  2. ask what you want to author, consume, validate, and distribute in small, plain-language batches;
  3. recommend a destination and starter configuration;
  4. show the complete proposed files and commands before writing; and
  5. scaffold and validate the new repository locally after you confirm it.

Creating files locally does not publish anything. Creating a GitHub repository, pushing commits, applying settings, or publishing a plugin, marketplace entry, or release requires separate approval.

This workflow is intentionally absent from the shared skill catalog: it depends on this commons repository's templates and conventions and is not installed with the plugin.

Consuming a skill

With the GitHub CLI (>= 2.90, needs gh auth login):

# Install one skill, pinned to a tag or SHA, into the repo's .agents/skills/.
gh skill install JeremyKuhne/agent-skills <skill> --pin vX.Y.Z

# Later, review upstream drift as a normal diff.
gh skill update <skill>

gh skill install installs only the named skill; it does not resolve the portfolio's metadata.requires relationships. Check the Requires column in the skills catalog and install every hard dependency at the same pin. For example, cswin32-com requires cswin32-interop:

gh skill install JeremyKuhne/agent-skills cswin32-interop --pin vX.Y.Z
gh skill install JeremyKuhne/agent-skills cswin32-com --pin vX.Y.Z

--pin records the exact version (and writes provenance: source repo, ref, and tree SHA into the copy's frontmatter). Commit the vendored copy plus any overlay.

As a plugin (skills + agent personas + MCP servers together):

copilot plugin marketplace add JeremyKuhne/agent-skills
copilot plugin install agent-skills@jeremykuhne-agent-skills

Installing the plugin brings in the MCP servers declared in .mcp.json. The client may start them on demand: the NuGet server runs locally via dnx and is version-pinned (NuGet.Mcp.Server@1.4.3), while microsoft-learn is a hosted HTTP endpoint on an official Microsoft domain. Review .mcp.json before installing if that matters in your environment.

The plugin install is smoke-tested with GitHub Copilot CLI. The skill cores use the vendor-neutral Agent Skills format and can be vendored for other compatible hosts, including Claude Code, but this repository does not currently claim a CI-tested Claude plugin-marketplace install.

Layout

PathPurpose
skills/The shared skill cores (the gh skill source).
agents/Portable agent personas (e.g. a generic reviewer).
plugin.jsonGitHub Copilot CLI plugin manifest (skills + agents + MCP).
.github/plugin/marketplace.jsonMarketplace listing so the plugin is installable by name.
.mcp.jsonMCP servers the skills rely on (microsoft-learn, NuGet).
.agents/skills/create-skill-repo/Repository-local workflow for creating a separate Agent Skills repository; not part of the shared catalog or plugin.
FORMAT.mdThe skill file format authors follow when contributing a core.
RELEASING.mdShared SemVer, release gates, and immutable-tag procedure.
CONTRIBUTING.mdHow to author and validate a shared skill core.
CODE_OF_CONDUCT.mdThe behavior expected of everyone in the project.
SECURITY.mdHow to report a vulnerability privately.
LICENSEThe MIT license this project is distributed under.

Governance

Consumers vendor pinned, provenance-stamped copies, so stability of published refs matters. The governance gate: a LICENSE (MIT), a security policy, contribution terms, and a code of conduct; --pinned installs; and immutable releases so a pinned tag cannot be rewritten after the fact.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados