skill-idea-miner

---
name: skill-idea-miner
description: Use when running the weekly skill-generation pipeline to mine session logs and score, dedupe, and backlog new skill idea candidates.
category: Strategy R&D
version: 1.0.0
tools: []
---

# Skill Idea Miner

Automatically extract skill idea candidates from coding-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.
- Manual backlog refresh.
- Dry-run to preview candidates without LLM scoring.

## Methodology

### Stage 1 — Session log mining
1. Enumerate session logs from allow-listed projects over the past 7 days (by file modification time, confirmed against a timestamp field in the log).
2. Extract external user messages and assistant tool-usage patterns from each session.
3. Run deterministic signal detection over that history:
   - Skill-usage frequency (references to existing skill paths).
   - Error patterns (non-zero exit codes, error flags, exception keywords).
   - Repetitive tool sequences (3 or more tools repeated 3 or more times).
   - Automation-request keywords, checked across languages.
   - Unresolved requests (a 5+ minute gap after a user message with no resolution).
4. Feed the detected signals to an LLM pass for idea abstraction — turning raw session evidence into a candidate title, description, and category.
5. Output the raw candidate list with its evidence trail.

### Stage 2 — Scoring and deduplication
1. Load the existing skill library's names and descriptions from their frontmatter.
2. Deduplicate new candidates against both existing skills and the current backlog using Jaccard similarity, with a threshold above 0.5 counting as a duplicate.
3. Score each non-duplicate candidate on three axes, each 0-100: Novelty (differentiation from existing skills), Feasibility (technical implementability), and Trading Value (practical value for investors/traders).
4. Compute a composite score as 0.3 x Novelty + 0.3 x Feasibility + 0.4 x Trading Value — trading value weighted highest since that's the practical bar for inclusion.
5. Merge scored, non-duplicate candidates into the prioritized backlog, each carrying its evidence, category, and score breakdown, with status `pending` until picked up by skill-designer.