SkillOpt
Microsoft research code that improves a skill document by scoring it, the way training adjusts model weights.
Install
npx skills add microsoft/SkillOpt -a claude-code
npx skills add microsoft/SkillOpt -a claude-code -g
Or copy the skill folder into .claude/skills/ (project) or ~/.claude/skills/ (all projects).
npx skills add microsoft/SkillOpt -a codex
npx skills add microsoft/SkillOpt -a codex -g
Or copy the skill folder into .agents/skills/ (project) or ~/.agents/skills/ (all projects).
npx skills add microsoft/SkillOpt -a openclaw
npx skills add microsoft/SkillOpt -a openclaw -g
Or copy the skill folder into skills/ (project) or ~/.openclaw/skills/ (all projects).
Clone the repository and follow the README to configure a backend and benchmark.
When to use
Use it when you have a skill with a measurable benchmark and want to improve it by score instead of by hand.
SkillOpt is Microsoft research code that improves a skill document the way training adjusts model weights, but without touching the model. It uses epochs, batch sizes, learning rates and validation gates. A separate optimizer model reads scored rollouts and proposes small add, delete or replace edits. By default an edit is kept only if it strictly improves a held-out validation score, and a learning-rate budget limits how much changes per step.
There's a web UI, and you can add your own backends and benchmarks. It's published on PyPI, backed by a paper that was accepted as a NeurIPS 2026 Spotlight, and documented in a versioned docs site and technical blog. It's research software, so expect to read the docs before running it.
Related skills
-
rea
Lets your agent reverse engineer an app from the outside, down to native binaries, and then build its own version.
-
superpowers
A set of development habits for your agent: test first, debug systematically, and plan before you code.
-
skills
The skills Matt Pocock uses in his own day-to-day engineering work.
-
ECC
A ready-made setup for coding agents that adds skills, persistent memory and security checks.