Agent Skills - Repositório de Skills Compartilhadas
JeremyKuhne/agent-skills
Skills portáveis compartilhadas para agentes em .NET com versionamento
Ver no GitHubMCP 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-skills1 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
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:
- explain how original skills are published and how other repositories use version-pinned copies;
- ask what you want to author, consume, validate, and distribute in small, plain-language batches;
- recommend a destination and starter configuration;
- show the complete proposed files and commands before writing; and
- 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
| Path | Purpose |
|---|---|
skills/ | The shared skill cores (the gh skill source). |
agents/ | Portable agent personas (e.g. a generic reviewer). |
plugin.json | GitHub Copilot CLI plugin manifest (skills + agents + MCP). |
.github/plugin/marketplace.json | Marketplace listing so the plugin is installable by name. |
.mcp.json | MCP 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.md | The skill file format authors follow when contributing a core. |
RELEASING.md | Shared SemVer, release gates, and immutable-tag procedure. |
CONTRIBUTING.md | How to author and validate a shared skill core. |
CODE_OF_CONDUCT.md | The behavior expected of everyone in the project. |
SECURITY.md | How to report a vulnerability privately. |
LICENSE | The 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.