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.
SkillOpt from Microsoft Research (released July 2026, already 16,000 stars) treats natural-language skill documents (SKILL.md files) as trainable parameters for LLM agents. Without modifying model weights, it runs an iterative refinement loop: the optimizer scores agent rollouts against benchmarks and makes bounded edits — add, delete, or replace fragments — accepting only changes that pass a held-out validation test. The result is a deployable best_skill.md artifact of 300–2,000 tokens that plugs directly into Claude Code, Codex, or any harness that reads skill files. Microsoft's own benchmarks show +19.1-point accuracy lift inside Claude Code, +24.8 inside Codex, and +23.5 in direct GPT-5.5 chat, across six benchmarks and seven models. Supports OpenAI, Azure, Claude, Qwen, and MiniMax backends; includes SkillOpt-Sleep mode for nightly offline self-evolution and a WebUI dashboard for monitoring.
Why a vibe-coder should care
If your agent keeps making the same category of mistakes, SkillOpt finds the pattern and rewrites the skill instruction until the errors disappear — without you tuning prompts by hand. A 20-point accuracy improvement directly means fewer re-runs and less time reviewing agent output.
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▌ More finds