Claude Skills — Conjunto de Skills para Claude Code
DN-OpenSource/claude-skills
Playbooks reutilizáveis para Claude Code — equipes, segurança, memória e governança
Ver no GitHubColeção de skills reutilizáveis que ensinam Claude Code a executar tarefas específicas de forma disciplinada e composável. Inclui skills para trabalho em equipe (teammates), edição segura de codebases (codebase-guardian), persistência de contexto (memory) e governança via contratos (agents-dox). Cada skill funciona independentemente e se compõe com as demais.
Instalação
npx convez-ai DN-OpenSource/claude-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
- Coordenar múltiplos agentes Claude em paralelo sem orquestrador centralizado
- Editar codebases existentes com validação contra a toolchain real
- Manter contexto e learnings persistentes entre sessões via MEMORY.md
- Estabelecer contratos de trabalho hierárquicos via AGENTS.md (DOX)
Veredito

Conjunto maduro e bem arquitetado para quem trabalha com Claude Code em projetos reais. A composição entre skills é elegante e o foco em segurança (codebase-guardian) e persistência (memory) resolve dores genuínas em fluxos multi-agente. Requer entendimento do ecossistema Claude Code, não é entry-level.
README
A tradução para o português deste README ainda está na fila.
claude-skills
A collection of custom skills for Claude Code. Each skill works on its own, and they compose with one another (see How the skills work together).
Version: 1.2.0 (all plugins)
What are skills?
Skills are Markdown files that teach Claude how to handle specific tasks — a multi-agent protocol, a document format, a repeatable workflow. Claude reads the relevant skill at the start of a task and follows its instructions. Think of them as reusable playbooks you install once and get forever.
Skills in this repo
| Skill | Version | Works in | Description |
|---|---|---|---|
| teammates | 1.2.0 | Claude Code | Run any task as a flat team of Claude Code subagents — no orchestrator, no hierarchy. Agents share a JSON manifest, claim work, message each other, and merge outputs. |
| codebase-guardian | 1.2.0 | Claude Code | A disciplined four-phase loop for editing existing codebases safely — validate against the real toolchain, match or deliberately change the existing pattern, trace the ripple, and record learnings to MEMORY.md. Covers TypeScript/Node, Python/Django, Flutter/Dart, and Rust/Tauri. |
| memory | 1.2.0 | Claude Code | Durable, hierarchical memory files so context survives across sessions — bootstrap, read, and update MEMORY.md at the repo root and per module, plus user-level USER.md. The persistence layer codebase-guardian builds on. |
| agents-dox | 1.2.0 | Claude Code | The DOX framework — a hierarchy of AGENTS.md files, each a binding work contract for its subtree. Read the root-to-nearest chain before editing; run a DOX pass to update the owning AGENTS.md and affected parents/children after. Bootstraps the tree if a repo has none. |
How the skills work together
Each skill is self-contained — install any one on its own and it works with no dependency on the others. They're also designed to compose:
- codebase-guardian + memory — guardian reads
MEMORY.mdbefore editing and writes learnings back after, so conventions and ripple traps accumulate across sessions. Withoutmemoryinstalled, guardian still maintainsMEMORY.mdfiles by hand; with it, that lifecycle is automatic. - teammates + memory — peers can read the shared
MEMORY.mdfor project context before claiming work, and the merge step can fold new learnings back in. - teammates + codebase-guardian — when a peer team is editing real code, each peer follows the guardian loop (validate with the toolchain, match the pattern, trace the ripple) for its own work item, so parallelism doesn't cost safety.
- agents-dox + everything — in a repo that uses the DOX
AGENTS.mdhierarchy, the contract chain governs the others: guardian reads the chain during Orient and folds the DOX pass into its closeout, and each teammates peer reads the same chain before claiming work.agents-dox(instructions/contracts inAGENTS.md) andmemory(durable facts inMEMORY.md) are complementary, not competing.
None of these are required: a skill never errors or stalls because a companion skill is absent.
Installation
Check the official Claude Code docs for the most current installation steps — the install mechanism may have been updated since this README was written.
Adding a new skill
-
Create a new directory under
skills/:skills/your-skill-name/ └── SKILL.md -
SKILL.mdmust start with YAML frontmatter:--- name: your-skill-name description: When to trigger this skill and what it does. Be specific — this is what Claude reads to decide whether to use the skill. --- # Your Skill Name ... -
Open a PR with a one-line summary of what the skill does and when to use it.
Structure
claude-skills/
├── README.md
└── skills/
├── teammates/
│ ├── README.md
│ └── SKILL.md
├── codebase-guardian/
│ ├── README.md
│ ├── SKILL.md
│ └── references/ ← per-stack commands & ripple traps
├── memory/
│ ├── README.md
│ └── SKILL.md
└── agents-dox/
├── README.md
└── SKILL.md
Each skill lives in its own subdirectory. If the skill needs supporting files (scripts, reference docs, templates), they go in subdirectories alongside SKILL.md:
skills/your-skill/
├── SKILL.md
├── scripts/ ← executable helpers
├── references/ ← docs loaded on demand
└── assets/ ← templates, fonts, etc.
License
Apache-2.0