All Tools
S
Fine-tuningFreeOpen Source
SKILLOPT
Train agent skills like neural nets, without touching weights
MIT
ABOUT
Agent skills are usually hand-written or one-shot generated, so they do not improve under feedback. SkillOpt treats the skill markdown as trainable state, applies bounded edits from rollouts, and only keeps changes that beat a held-out validation score, with zero extra model calls at deploy time.
INSTALL
pip install skilloptINTEGRATION GUIDE
1. Optimize a Claude Code or Codex skill from scored trajectories with a validation gate
2. Train a reusable best_skill.md that transfers across GPT, Claude, and Qwen agents
3. Run SkillOpt-Sleep overnight to harvest sessions and consolidate validated skills
4. Benchmark skill optimization on SearchQA or other SkillOpt environments
TAGS
pythonagentsskillsmicrosoftoptimizationfrozen-llm