SkillOpt Project Page
Official Microsoft project site for SkillOpt — a text-space optimizer that trains reusable natural language skills for frozen language model agents.
Design principles
- Skill as parameter: The markdown skill document is the optimization target; target model, backend, and harness stay fixed
- Validation-gated edits: Plausible-sounding text changes are rejected unless held-out validation improves
- Deep-learning discipline: Rollout evidence (forward pass), reflection (backward pass), bounded textual learning rate
- Deployment efficiency: No extra inference calls; portable
best_skill.mdartifact
Key components
| Component | Role |
|---|---|
| Optimizer model | Proposes structured skill edits from scored rollouts |
| Validation gate | Strict held-out performance check before accepting edits |
| Edit budget | Textual learning rate controlling per-step change magnitude |
| Rejected-edit buffer | Negative memory preventing repeated failed edits |
| Slow/meta update | Epoch-wise comparison stabilizes long training runs |
Transfer experiments
Optimized skills retain value when moved across:
- Model scales (frontier → smaller models)
- Execution harnesses (Codex CLI ↔ Claude Code)
- Related benchmarks without re-optimization
Resources
- Paper: arXiv:2605.23904
- Code: github.com/microsoft/SkillOpt
- PyPI:
pip install skillopt
Authors
Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, Chong Luo (Microsoft Research Asia and academic collaborators).