SkillOpt: Microsoft Trains Agent Instructions Instead of Model Weights, Gains +23% Accuracy

Microsoft’s SkillOpt optimizes natural-language skill documents for AI agents through automated edits and validation gates, boosting GPT-5.5 accuracy by +23.5% with zero added inference cost.
artificial-intelligence
Author

Kabui, Charles

Published

2026-06-03

Keywords

skillopt, agent-skills, text-space-optimization, microsoft-research, prompt-optimization

Download as PDF

Microsoft Research released SkillOpt, a text-space optimizer that trains natural-language skill documents instead of model weights. A separate optimizer model runs the frozen target agent on scored batches, reflects on failures and successes separately, then proposes bounded add/delete/replace edits to the skill document. An edit is accepted only if it improves a held-out validation score. The approach borrows deep learning concepts (epochs, learning rates, minibatches) but applies them to text. Across six benchmarks, seven target models, and three execution harnesses (direct chat, OpenAI Codex, and Claude Code), SkillOpt was best or tied-best in all 52 evaluated settings. On GPT-5.5, it lifted average accuracy by +23.5% in direct chat, +24.8% inside Codex, and +19.1% inside Claude Code versus running the same models with no skill. The final artifact is a compact best_skill.md file, typically 300 to 2,000 tokens long.

A team can boost agent performance by sharing a text file rather than retraining a model. Skills transfer across model scales and execution environments: a skill optimized on GPT-5.5 in Codex still works when moved to Claude Code or to a nearby math benchmark without further optimization. Since the skill is just context tokens at runtime, there’s zero added latency or compute cost. The system is open-source under MIT license and installable via pip install skillopt.

This fits a growing pattern alongside Alibaba’s SkillClaw, Anthropic’s CLAUDE.md files, and VS Code’s .instructions.md: structured text instructions are becoming first-class optimization targets. Training the prompt may become as routine as training the model.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {SkillOpt: {Microsoft} {Trains} {Agent} {Instructions}
    {Instead} of {Model} {Weights,} {Gains} +23\% {Accuracy}},
  date = {2026-06-03},
  url = {https://toknow.ai/posts/microsoft-skillopt-text-space-optimizer-agent-skills-zero-inference-overhead/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “SkillOpt: Microsoft Trains Agent Instructions Instead of Model Weights, Gains +23% Accuracy.” https://toknow.ai/posts/microsoft-skillopt-text-space-optimizer-agent-skills-zero-inference-overhead/.