Runs at project start, middle, or end to eliminate permission prompts and organize your Claude Code setup. Spins up parallel sub-agents to audit permissions (leaked secrets, missing MCP servers, stale entries, legacy syntax) and context quality (oversized files, stale references, missing documentation). For new projects it detects your stack from files present and generates settings.local.json with the right presets, plus a starter CLAUDE.md. The sub-agent architecture is smart: keeps heavy filesystem scanning out of your main conversation so you get clean summaries instead of walls of diagnostic output. Also handles mid-project cleanup and end-of-session learning capture. Basically prevents the configuration drift that makes Claude Code chatty.
npx skills add https://github.com/jezweb/claude-skills --skill project-health