Microsoft SkillOpt Trains Markdown Skills for Frozen LLMs. 52/52 Benchmark Wins. Your CLAUDE.md Can Be Optimized Like a Neural Network.

Microsoft Research dropped a paper that should change how every builder thinks about their agent configuration files. SkillOpt (arXiv 2605.23904) treats a Markdown document as an external parameter of a frozen LLM and applies learning rate, batch, and momentum concepts in text space. Not metaphorically. Literally. They train your skill file.

The results are absurd. Across 6 benchmarks, 7 models, and 3 execution harnesses (direct chat, Codex, Claude Code), SkillOpt wins or ties on all 52 evaluated cells. On GPT-5.5: +23.5 points in direct chat, +24.8 in Codex, +19.1 in Claude Code. Those aren't marginal improvements. That's the difference between a mediocre agent and a good one.

I maintain a CLAUDE.md file for every project I work on. Most builders I know do the same. We've been treating prompt engineering as intuition, taste, trial and error. SkillOpt says: no, this is an optimization problem with a tractable solution. You can systematically search the space of possible skill file configurations and find ones that measurably improve your model's task performance.

The mechanism: SkillOpt iteratively edits sections of the Markdown skill file, evaluates performance on a held-out task set, and applies something analogous to gradient descent in text space. The "learning rate" controls how aggressively edits are made. The "momentum" preserves successful patterns across iterations. It's MIT licensed and open source.

This connects directly to the cost story. If Copilot is now charging you per token, and SkillOpt can get you better results in fewer attempts (because the model understands your intent more precisely on the first try), then optimized skill files aren't just about quality. They're about economics. Better skills = fewer wasted tokens = lower bills.

What builders should do: if you maintain CLAUDE.md, AGENTS.md, .cursor/rules, or any agent skill files, this is worth running against your actual task distribution. The assumption that "prompt engineering is art not science" just took a hit. It's still art. But now there's a gradient.