recursive-improve

Run `npx skills add kayba-ai/recursive-improve --all -g -y` in your terminal to install all skills in this suite globally.

Add two lines of tracing code to capture your agent's runs, then invoke the /recursive-improve skill. It analyzes the traces, finds failure patterns, and applies targeted code and prompt fixes on a separate branch.

Yes. The /ratchet skill runs an autonomous loop that improves your agent, evaluates the results, and keeps or reverts each change. You review a summary of what improved when you return.

Yes. The skills run directly inside Claude Code or Codex, and the tracing library captures calls from OpenAI, Anthropic, and LiteLLM clients automatically.

Run the /benchmark skill to measure metric quality before and after each change. The dashboard visualizes every improvement cycle with before-and-after metrics so you can merge wins and discard regressions.