Skill Idea Miner

Runs weekly to scan your Claude Code session logs and extract potential skill ideas based on what you've actually been doing. It looks for patterns like repeated tool sequences, error clustering, and automation requests, then scores candidates on novelty, feasibility, and trading value before adding them to a backlog. The deduplication uses Jaccard similarity to avoid reinventing existing skills. Honest take: this is meta tooling for people building Claude skills at scale, not something you'd use unless you're running a regular skill generation pipeline. The dry run mode is handy for previewing what it would extract before burning LLM calls on scoring.