Skill Improver

A scheduled audit loop. Each run pulls newly finished agent conversations across configured projects, identifies steering moments, attributes each one to a specific skill (or the orchestration around it), and ships a single PR with all proposed skill edits — with the why spelled out in both the PR and the conversation that triggered the run.

The goal is continuous, evidence-based skill improvement. Every time something goes wrong, the question is: what went wrong, and who should have caught it? If a skill could have prevented it, the skill changes.

Required tools

• git and gh (GitHub CLI, authenticated) — for branch + PR creation
• python3 — runs the conversation puller and the brief discovery script
• Read access to ~/.codex/sessions/, ~/.codex/archived_sessions/, and ~/.claude/projects/
• Read access to the .briefs/ directory inside each configured project root (and its worktrees) — this is where the brief skill writes its HTML artifacts

If any are missing, stop and ask.

Configuration

config.json (next to this file) lists projects to analyze and the path patterns where their conversations live (main repo, codex worktrees, claude worktrees, cursor worktrees, gwt-worktrees). To add a project, append a new entry: