SkillOpt — Microsoft Open Source Repository

MIT-licensed framework from Microsoft Research for optimizing AI agent skill markdown documents without modifying underlying model weights.

Core concept

Agent skills are stored as .md files containing procedural knowledge — domain heuristics, tool-use policies, output constraints, failure modes. SkillOpt treats these documents as trainable external state of a frozen LLM agent.

Optimization loop

  1. Run frozen target model on scored task batches
  2. Separate optimizer model proposes bounded add/delete/replace edits
  3. Held-out validation gate accepts edits only when performance strictly improves
  4. Textual learning-rate budget limits edit magnitude per step
  5. Rejected-edit buffer prevents repeated failed modifications
  6. Epoch-wise slow/meta update for long-horizon stability

Output artifact

Exports compact best_skill.md (~300–2,000 tokens, median ~920) with zero additional inference-time model calls at deployment.

Quick start

pip install skillopt
python scripts/train.py --config configs/searchqa/default.yaml \
  --split_dir /path/to/split --optimizer_model gpt-5.5 --target_model gpt-5.5

Benchmark results (paper)

  • Best or tied on all 52 (model, benchmark, harness) cells evaluated
  • GPT-5.5: +23.5 average absolute points vs no-skill baseline (direct chat)
  • Cross-harness transfer: Codex-trained spreadsheet skill → +59.7 points in Claude Code
  • Training cost: ~$1–5 per single-task skill via community frameworks (GBrain)

License

MIT — compatible with commercial deployment and modification.