# SkillOpt

> MIT skills optimizer — trains skill.md as params for frozen agents

**Category:** AI Agents  
**Pricing:** open-source  
**URL:** https://microsoft.github.io/SkillOpt/docs/  
**GitHub:** https://github.com/microsoft/SkillOpt  
**Added:** 2026-09-30

---

## The Hook

SkillOpt turns agent skill documents into trainable parameters — rollout, reflect, gated update — without touching model weights or renting a skills SaaS.

## What's Good

MIT (microsoft/SkillOpt). `pip install skillopt` (PyPI) plus research checkout path. Treats `best_skill.md` as the adapter: trajectory-driven edits, textual learning-rate budget, validation gate, deployable artifact. BYO LLM backends (OpenAI/Azure/Claude/Qwen/…). SkillOpt-Sleep companion for offline session harvest. Not a paid wall — bring your own keys and verifiers.

## The Catch

★17,883 in about 70 weeks (~255 a week): judge it on the optimizer loop, not the star count. Vs Superpowers / agent-skills (Addy) / Matt Pocock skills packs: those ship hand-authored SKILL.md workflows and session methodology; SkillOpt trains/evolves skills as parameters for frozen agents. Training burns provider tokens; plugin shells for Claude/Codex/etc. live in-repo more than in the PyPI wheel.

## Verdict

MIT skill trainer when frozen agents need evolved skill.md — not another hand-pack; ignore the star spike.