Skill Idea Miner

Automatically extract skill idea candidates from Claude Code session logs, score them for novelty, feasibility, and trading value, and maintain a prioritized backlog for downstream skill generation.

When to Use

• Weekly automated pipeline run (Saturday 06:00 via launchd)
• Manual backlog refresh: python3 scripts/run_skill_generation_pipeline.py --mode weekly
• Dry-run to preview candidates without LLM scoring

Prerequisites

• Python 3.10+ with pyyaml package
• Claude CLI installed and authenticated (claude --version to verify)
• Session logs in ~/.claude/projects/<project>/ (created automatically by Claude Code)
• No API keys required (uses Claude CLI for LLM calls)

Workflow

Quick Start

# Dry-run: preview mined candidates without LLM scoring
python3 scripts/mine_session_logs.py --dry-run --output-dir reports/

# Full mining with scoring (requires Claude CLI)
python3 scripts/mine_session_logs.py --output-dir reports/

# Score existing candidates
python3 scripts/score_ideas.py \
  --candidates reports/raw_candidates.yaml \
  --output-dir logs/

Stage 1: Session Log Mining

1. Enumerate session logs from allowlist projects in ~/.claude/projects/
2. Filter to past 7 days by file mtime, confirm with timestamp field
3. Extract user messages (type: "user", userType: "external")
4. Extract tool usage patterns from assistant messages
5. Run deterministic signal detection:
   • Skill usage frequency (skills/*/ path references)
   • Error patterns (non-zero exit codes, is_error flags, exception keywords)
   • Repetitive tool sequences (3+ tools repeated 3+ times)
   • Automation request keywords (English and Japanese)
   • Unresolved requests (5+ minute gap after user message)
6. Invoke Claude CLI headless for idea abstraction
7. Output raw_candidates.yaml

Stage 2: Scoring and Deduplication