microsoft/SkillOpt

A Microsoft research tool that automatically optimizes agent skills through a training loop — rollout, reflect, aggregate, select, update, evaluate — with built-in benchmarks.

13k stars 1.2k forksLicense: MITUpdated 2026-07-17
Researchoptimizationtrainingbenchmarksresearch
View on GitHub

Why it matters

Most skills are written and tuned by hand, which caps how good they get. SkillOpt treats a skill like something you can measurably improve, iterating against benchmarks across multiple model backends and keeping the version that scores best. For anyone serious about skill quality rather than vibes, it turns optimization into a repeatable, evaluated process.

Use cases

  • Automatically iterate and improve a skill against benchmarks
  • Compare skill performance across OpenAI, Claude, and other backends
  • Run a structured rollout-reflect-evaluate optimization loop
  • Ground skill quality in measured results instead of intuition

Quickstart

pip install skillopt

Related repositories

virgiliojr94Research

book-to-skill

A skill (and companion CLI) that converts a book or long document into a reusable agent skill, distilling its methods into something Claude can apply.

8.7k 1.1k
booksknowledgeextractionskill-authoring
joeseesunResearch

qiaomu-anything-to-notebooklm

A skill that converts assorted sources — articles, documents, media — into clean material ready to load into Google NotebookLM as a knowledge base.

5.6k 616
notebooklmresearchsourcesknowledge

Maintain this repo? Submit an update.

Submit an update