Claude Code Skills — Comandos para Documentação Automática
pearl04/claude-skills
Suas sessões de código se documentam automaticamente
Ver no GitHubConjunto de slash commands para Claude Code que automatiza a documentação de sessões de coding. Gera automaticamente logs detalhados (devlog.md) e notas de handoff (SESSION_CONTEXT.md) ao final de cada sessão, eliminando a necessidade de documentação manual e preservando contexto do projeto.
Instalação
npx convez-ai pearl04/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
- Manter histórico estruturado de decisões técnicas e pivôs durante desenvolvimento
- Facilitar handoffs entre desenvolvedores com notas de contexto automáticas
- Reconstruir rapidamente o estado do projeto sem perder 20 minutos de contexto
- Criar base de conhecimento durável e pesquisável sobre evolução do projeto
Veredito

Solução pragmática que resolve um problema real de desenvolvedores: documentação que morre no esquecimento. Abordagem minimalista (markdown + append) é mais robusta que alternativas com embeddings ou vector stores. Utilitário direto, sem bloat, que se integra naturalmente ao fluxo de trabalho existente.
README
A tradução para o português deste README ainda está na fila.
claude-skills
Your coding sessions document themselves.
A growing collection of Claude Code slash commands. The first release: a three-command system that turns every coding session into structured project documentation, automatically.
The problem
Project docs die. Handoffs are painful. You ship a feature, move on, and two weeks later you can't remember why you built it the way you did. You sit down to code and waste 20 minutes rebuilding context.
The fix: /end-session
One command at the end of every Claude Code session. It reads the entire conversation and writes two things to your project:
devlog.md— a detailed, dated entry covering goals, decisions, what you built, pivots, mistakes, learnings, and next steps. Permanent. Searchable. This is your project documentation, accumulating session by session.SESSION_CONTEXT.md— a short handoff note (status, last worked on, next up). Rolling window of the last 5 sessions. Read it at the start of the next session and pick up exactly where you left off.
Zero manual writing. The LLM already has all the context — this just captures it.
The three commands
| Command | What it does |
|---|---|
/end-session | Runs /session-context + /wrap-up back-to-back, then offers a git push |
/session-context | Prepends a brief handoff note to SESSION_CONTEXT.md |
/wrap-up | Prepends a detailed project log entry to devlog.md |
You only ever run /end-session. The other two exist so you can invoke either one standalone if you want.
Install
Copy the three files into your Claude Code commands directory:
git clone https://github.com/pearl04/claude-skills.git
cp claude-skills/commands/*.md ~/.claude/commands/
Or download just the files you want from commands/ and drop them in ~/.claude/commands/.
Restart Claude Code (or start a new session) and type /end-session to try it.
How it looks
After a few sessions, your project root has:
your-project/
├── devlog.md ← full chronological history (forever)
├── SESSION_CONTEXT.md ← last 5 handoff notes
└── ...your code
devlog.md entry example:
## Session: 2026-04-12
### Goal
- Fix the fabrication problem in the generation pipeline
### Built
- backend/pipeline/question_generator.py (NEW)
- backend/api/routes/generation.py — added /prepare endpoint
- ...
### Key Decisions
- Follow-up questions BEFORE generation, not after
- Anti-fabrication rule as FIRST paragraph of system prompt
### Next Steps
1. Wire frontend to use /prepare
2. Test edit → style correction learning
Why it works
Most "AI memory" systems try to solve context with embeddings or vector stores. This one is radically simpler: append structured markdown to the project root after every session. The LLM reads it next time. That's it.
The best docs are the ones you don't have to write.
More commands
The repo has grown beyond the session-documentation trio. These commands are general-purpose, stack-agnostic, and contain no personal configuration:
| Command | What it does |
|---|---|
/evaluate-skill | Audits a Claude Code skill/command against a 10-point best-practices rubric |
/security-review | Validates external links, repos, and packages for supply-chain risk before you act |
/qa | Detects your stack, writes targeted tests for changed files, runs them, reports back |
/add-feature | Quick-captures a feature into BACKLOG.md with Impact × Ease scoring |
/review-backlog | Triage session for your project's BACKLOG.md |
/create-doc | Generates comprehensive project documentation (PROJECT_DOC.md) |
/gap-analysis | Analyses any content (URL / file / image) and surfaces gaps across your projects |
Drop any of them into ~/.claude/commands/ (still supported) or ~/.claude/skills/ and start a new session.
License
MIT. Use it, fork it, remix it.