GitHub radar

SkillOpt by Microsoft: Train Reusable LLM Agent Skills

SkillOpt from Microsoft Research automatically optimizes natural-language skill documents for LLM agents — delivering a 19–25 point accuracy lift inside Claude Code, Codex, and GPT-5.5.

01microsoft/SkillOpt 16kPython

SkillOpt from Microsoft Research (July 2026, 16k Stars) lets you train natural language descriptions of skills (SKILL.md files) for LLM agents. It does this using an iterative optimization process where an optimizer evaluates agent rollouts on benchmark tasks and applies small modifications (additions, deletions, substitutions) to the skills while ensuring these changes satisfy a holdout validation set. At the end of the optimization process, you get a final best_skill.md file with 300–2,000 tokens which can be used with Claude Code, Codex, or any other skill harness. On their benchmarks, they see an improvement of +19.1 points when using this method in Claude Code, +24.8 points in Codex, and +23.5 points in direct GPT-5.5 chat conversations. They support OpenAI, Azure, Claude, Qwen, and MiniMax backends. They also have SkillOpt-Sleep which allows your agent to evolve overnight by itself, and a WebUI interface to track its progress.

Why a vibe-coder should care

If your agent tends to make similar types of errors, SkillOpt will notice this and modify the skill description accordingly so that your agent stops making those errors. This means that you don’t have to manually tweak prompts! An increase in accuracy of 20 points means that you will have to rerun your agent less often and spend less time checking its results.

How to install

Copy this and send it to your agent — Claude Code, Codex, any of them:

Use SkillOpt from here: https://github.com/microsoft/SkillOpt — install via `pip install skillopt`, run the optimizer on the target SKILL.md file, generate best_skill.md and show me what changed. I have a model API key, just ask.

Runs on a regular laptop — Python 3.10 and a language model API key required.

Open on GitHub