Skill Idea Miner reads Claude Code session logs from the last seven days, looks for patterns such as repeated tool sequences, error messages, automation requests and unanswered questions, and asks the Claude CLI to turn them into skill ideas.
A second stage removes near-duplicates of existing skills and backlog entries using Jaccard similarity, scores the rest on novelty, feasibility and trading value, and merges them into a YAML backlog. It is built for the weekly skill generation pipeline in this repository.
What You Get
- A raw_candidates.yaml file with titles, evidence and categories
- A scored YAML backlog with novelty, feasibility, trading value and a composite score
- A dry-run mode that previews candidates without calling the model
- A scoring rubric reference for the signal detection criteria
Common Use Cases
- Running the weekly skill generation pipeline on a schedule
- Refreshing the idea backlog by hand after a busy week of sessions
- Checking which recurring pain points have no skill yet
What Makes It Unique
It derives ideas from how you used Claude Code that week, running deterministic signal detection first and the model second, and weights trading value highest in the composite score (40 percent).
Under The Hood
Python scripts read session log files under the Claude projects folder in your home directory, limited to an allowlist of projects and to the last seven days, and read existing SKILL.md frontmatter. They call the Claude CLI in headless mode as a subprocess for idea abstraction and scoring, and write YAML files into a reports folder and a logs folder. The SKILL.md also lists shell commands to run these scripts.