Skill idea miner

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

1. Load existing skills from skills/*/SKILL.md frontmatter
2. Deduplicate via Jaccard similarity (threshold > 0.5) against:
   • Existing skill names and descriptions
   • Existing backlog ideas
3. Score non-duplicate candidates with Claude CLI:
   • Novelty (0-100): differentiation from existing skills
   • Feasibility (0-100): technical implementability
   • Trading Value (0-100): practical value for investors/traders
   • Composite = 0.3 * Novelty + 0.3 * Feasibility + 0.4 * Trading Value
4. Merge scored candidates into logs/.skill_generation_backlog.yaml

Output Format

raw_candidates.yaml

generated_at_utc: "2026-03-08T06:00:00Z"
period: {from: "2026-03-01", to: "2026-03-07"}
projects_scanned: ["claude-trading-skills"]
sessions_scanned: 12
candidates:
  - id: "raw_2026w10_001"
    title: "Earnings Whispers Image Parser"
    source_project: "claude-trading-skills"
    evidence:
      user_requests: ["Extract earnings dates from screenshot"]
      pain_points: ["Manual image reading"]
      frequency: 3
    raw_description: "Parse Earnings Whispers screenshots to extract dates."
    category: "data-extraction"

Backlog (logs/.skill_generation_backlog.yaml)

updated_at_utc: "2026-03-08T06:15:00Z"
ideas:
  - id: "idea_2026w10_001"
    title: "Earnings Whispers Image Parser"
    description: "Skill that parses Earnings Whispers screenshots..."
    category: "data-extraction"
    scores: {novelty: 75, feasibility: 60, trading_value: 80, composite: 73}
    status: "pending"

Resources

• references/idea_extraction_rubric.md — Signal detection criteria and scoring rubric
• scripts/mine_session_logs.py — Session log parser
• scripts/score_ideas.py — Scorer and deduplicator