Claude Skills - Coleção de Skills Modulares
basje01/claude-skills
Monorepo de skills modulares e versionados para Claude Code com garantias de segurança
Ver no GitHubRepositório monorepo de skills reutilizáveis para Claude Code, cada um em seu próprio subdiretório. Facilita o carregamento automático de padrões operacionais, scripts de refresh e integrações com camadas de memória MCP quando condições de ativação são atendidas.
Instalação
npx convez-ai basje01/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
- Organizar e versionar skills personalizados para Claude Code
- Implementar rituals de refresh de dados externos com auditoria de conteúdo T5
- Estruturar integrações MCP (como camadas de memória pessoal) com segurança
- Manter histórico append-only de capturas de dados para rastreabilidade
Veredito

Framework sólido e opinado para gerenciar skills de Claude Code em escala, com ênfase em safety e rastreabilidade de conteúdo externo. Ideal para operadores que precisam de patterns reutilizáveis versionados e auditáveis.
README
A tradução para o português deste README ainda está na fila.
claude-skills
A continuously-updated collection of Claude Code skills authored by the operator. Each skill is a self-contained subfolder loaded by Claude Code (or any Claude harness honoring the same convention) when its trigger conditions match — so future sessions don't have to re-derive operational patterns from upstream docs every time.
Why a monorepo
Single source of truth, one git history for skill evolution, one CI/lint surface, and a coherent place to land cross-cutting refactors (e.g., when the external-AI-output safety rules tighten, every skill's T5 handling can be updated in one PR).
Each skill is isolated in its own subfolder. The Claude Code loader treats
~/.claude/skills/<name>/SKILL.md as the entry point; this repo populates
those entries via symlinks (see Local install below). The
loader follows symlinks — verified against the pre-existing
team_members/braid-reasoning symlink pattern.
Layout
claude-skills/
├── README.md ← this file
├── gbrain/ ← Garry Tan's personal-knowledge MCP memory layer
│ ├── SKILL.md ← auto-loaded body
│ ├── refresh.sh ← weekly grok-p refresh ritual
│ ├── com.bas.gbrain-skill-refresh.plist
│ ├── install-refresh-launchd.sh
│ ├── references/
│ │ ├── upstream-schema-excerpt.md
│ │ ├── refresh-procedure.md
│ │ ├── t5-grok-intel-<UTC>.md ← current capture
│ │ └── archive/<YYYY>/<MM>/ ← append-only history of captures
│ ├── README.md ← skill-specific docs
│ └── .gitignore ← launchd noise (.log files)
└── <future-skill>/ ← same pattern for any new skill
Local install
After cloning this repo (git clone git@github.com:basje01/claude-skills.git
into ~/Code/claude-skills/ to match the convention with the operator's other
repos), register each skill with the Claude Code loader by symlinking:
ln -s ~/Code/claude-skills/gbrain ~/.claude/skills/gbrain
The loader picks up the new entry on next session start.
Per-skill conventions
Each skill subfolder is expected to have:
SKILL.md— the loaded body. YAML frontmatter must includename,description, andmetadata.triggers. Body is markdown.README.md— skill-specific documentation: layout, cadence, safety contract.references/— longer-form supporting docs theSKILL.mdcan point at (T1 anchors, playbooks, captures)..gitignore— exclude transient artifacts (logs, runtime caches).
Anything skill-specific (refresh scripts, plists, installers) lives inside the skill's own subfolder, not at the repo root.
Safety model for skills that capture external AI output
Skills that surface T5 content (e.g., a refresh ritual that pulls intel from grok-p, openai, gemini, etc.) MUST:
- Capture verbatim text into a dated file under the skill's
references/, wrapped in an<external_ai_output source="<tool>" trusted="false">envelope. - Diff against the prior capture for operator review.
- Archive prior captures, never delete them — append-only timeline.
- Never auto-promote T5 claims into the SKILL.md body. Promotion requires the operator to verify the cited source independently (URL, commit sha, official docs).
- Auto-promote ONLY T1 facts (verified — repo HEAD, host config files, official docs).
This matches the global rule at ~/.claude/rules/external-ai-output.md.
The gbrain/refresh.sh here is the reference implementation; new skills
adopting an external-AI capture pattern should mirror its structure.
Cadence
Each skill is responsible for its own refresh cadence. The gbrain skill
uses a weekly launchd timer (Sunday 17:00 local), paired with a host-side
systemd timer that upgrades the gbrain service itself overnight (Sunday
03:17 UTC) so the host advances first, the skill captures the resulting
state in the evening. Skills with no remote content drift (pure reference
skills) need no refresh.
Related operator repos
bas-personal-bot— Hermes gateway + hardening + ops runbooks (the consumers of the gbrain skill)hyperliquid-prediction— the trading bot whose memory layer is gbrain
Adding a new skill
cd ~/Code/claude-skills
mkdir -p <new-skill>/references
# Write <new-skill>/SKILL.md with frontmatter (name, description, triggers)
# Write <new-skill>/README.md with layout + cadence + safety contract
# Add any refresh/install scripts inside <new-skill>/
git add <new-skill>
git commit -m "feat(<new-skill>): initial commit"
git push
ln -s ~/Code/claude-skills/<new-skill> ~/.claude/skills/<new-skill>
That's the entire registration. The Claude Code loader picks it up on the next session start.