SkillOpt

This page may contain stale information. Last updated: 2026-06-12

Overview

SkillOpt is an MIT-licensed open-source framework from microsoft Research Asia that optimizes AI agent skill markdown documents via validation-gated text-space training — without modifying underlying model weights.

Key Capabilities

• Treats .md skill files as trainable external state of frozen LLMs
• Propose-and-test loop with held-out validation gates
• Exports compact best_skill.md (~920 tokens median, max 2,000)
• Zero additional inference-time model calls at deployment

Benchmark Results (Microsoft Research)

• +23.5 average absolute points on GPT-5.5 vs no-skill baseline
• Best or tied on all 52 (model, benchmark, harness) cells
• Cross-harness transfer: Codex → Claude Code (+59.7 on spreadsheet skill)
• SpreadsheetBench: 41.8% → 80.7% on GPT-5.5
• Training cost: ~$1–5 per single-task skill (GBrain community framework)

Release

• 2026-06-11: Open-source release (MIT)
• GitHub: github.com/microsoft/SkillOpt
• Paper: arXiv:2605.23904
• PyPI: pip install skillopt

Related

• microsoft-skillopt-open-source

• microsoft

• agent-skills

• text-space-optimization

• skill-engineering

• enterprise-ai-reliability

• open-source-ai

Sources

• 2026-06-11-skillopt-github-microsoft
• 2026-06-11-skillopt-microsoft-project-page
• 2026-06-11-skillopt-arxiv-2605-23904
• 2026-06-11-microsoft-skillopt-open-source