convez.ai
SkillProdutividadeIntermediário
longlo8061

Skill Workshop

longlo8061/skill-workshop

Minerador de padrões em sessões Claude Code que gera skills reutilizáveis

Ver no GitHub
InicialPluginScore · de 0 a 100

Analisa o histórico de sessões do Claude Code para encontrar padrões recorrentes e gerar rascunhos de skills reutilizáveis. Identifica explicações repetidas, sequências de ferramentas e padrões de erro-correção, ranqueando candidatos por relevância para transformar em skills documentadas.

Instalação

$npx convez-ai longlo8061/skill-workshop

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
28%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Extrair workflows repetitivos em projetos e documentá-los como skills
  • Identificar correções frequentes e criar skills para contornar problemas conhecidos
  • Capturar explicações recorrentes sobre configurações específicas do ambiente
  • Automatizar a descoberta de oportunidades de reutilização de código

Veredito

Ferramenta inteligente para quem trabalha frequentemente com Claude Code e quer reutilizar padrões descobertos automaticamente. Excelente para documentar conhecimento tácito disperso em sessões, mas requer histórico de sessões ativo e ambiente Claude Code configurado.

README

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

Skill Workshop

Analyze your Claude Code session history to discover patterns that should become reusable skills.

Skill Workshop mines JSONL session files, finds repeating explanations, tool-chain workflows, and error→workaround patterns, then proposes skill candidates ranked by score. You pick the winners — it generates ready-to-use SKILL.md drafts.

What it finds

Signal typeWhat it catchesExample
Repeated explanationSame concept explained to Claude in 3+ sessions"Always use --break-system-packages with pip on this machine"
Tool chainSame sequence of tools used across sessionsGrep → Read → Edit → Bash for a specific workflow
WorkaroundErrors followed by user correctionsClaude keeps using wrong flags, user keeps fixing it

Installation

Copy the contents into your Claude Code configuration:

~/.claude/
├── agents/
│   └── session-analyzer.md    # Haiku-based extraction agent
└── skills/
    └── skill-workshop/
        └── SKILL.md            # Orchestrator skill

Or symlink from this repo:

ln -s /path/to/skill-workshop/agents/session-analyzer.md ~/.claude/agents/session-analyzer.md
mkdir -p ~/.claude/skills/skill-workshop
ln -s /path/to/skill-workshop/skills/skill-workshop/SKILL.md ~/.claude/skills/skill-workshop/SKILL.md

Usage

In any Claude Code session, run:

/skill-workshop

Or with an explicit project path:

/skill-workshop /path/to/project

The skill will:

  1. Find session files for the target project
  2. Delegate extraction to the session-analyzer agent (runs on Haiku for speed/cost)
  3. Present ranked candidates with evidence quotes
  4. Ask which candidates to develop into skill drafts
  5. Generate SKILL.md files in .claude/skills/

How it works

Phase 1 — Extract & Analyze (session-analyzer agent):

  • Parses JSONL session files using bash + jq/python3
  • Extracts user messages, tool sequences, and error→correction pairs
  • Groups by semantic similarity, scores candidates (frequency × type_weight)
  • Writes results to /tmp/skill-workshop-results.json

Phase 2 — Present & Generate (skill-workshop orchestrator):

  • Reads analysis results, presents candidates in a table
  • For approved candidates, generates proper SKILL.md files following skill authoring best practices
  • Outputs skills categorized as knowledge, workflow, or gotcha

Prerequisites

  • Claude Code with skills and agents support
  • Session history for the target project (Claude Code stores sessions in ~/.claude/projects/)

Tip: Claude Code deletes sessions after 30 days by default. To keep longer history, add "cleanupPeriodDays": 100000 to ~/.claude/settings.json.

Output

Results are written to /tmp/skill-workshop-results.json — a structured JSON with:

  • Project metadata and date range
  • Extraction stats (messages parsed, tool calls, errors found)
  • Up to 15 candidates ranked by score (0.0–1.0), each with evidence quotes

License

MIT

Ver o original no GitHub

Bom para

Relacionados