SkillOpt

Run `npx skills add microsoft/SkillOpt --all -g -y` in your terminal to install all skills in this suite globally.

It automatically improves your AI agent's skill files by testing proposed edits against held-out validation tasks, keeping only changes that measurably improve accuracy. No model retraining or weight changes are involved.

It is an offline self-improvement mode that reviews your past agent sessions, replays recurring tasks, and stages validated skill improvements for your review. Nothing changes in your live files until you explicitly adopt a proposal.

It ships integrations for Claude Code, Codex, Cursor, GitHub Copilot, and Devin, plus a reference adaptation for OpenClaw. It also supports multiple model backends including OpenAI, Azure, Claude, and Qwen.

The default mock backend is fully deterministic and makes no paid API calls, so you can test everything free. Real optimization runs use your chosen provider's API budget, with limits you control.