Skip to content

README.md

Latest commit

 

History

History
86 lines (61 loc) · 3 KB

README.md

File metadata and controls

86 lines (61 loc) · 3 KB

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 type What it catches Example
Repeated explanation Same concept explained to Claude in 3+ sessions "Always use --break-system-packages with pip on this machine"
Tool chain Same sequence of tools used across sessions Grep → Read → Edit → Bash for a specific workflow
Workaround Errors followed by user corrections Claude 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