Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses

SkillOpt comes out of a collaboration between Microsoft and three Chinese universities, and its pitch is refreshingly unglamorous: instead of fine-tuning a model, you optimize a single markdown file. The target model never changes. A separate optimizer model reads scored attempts and proposes small edits — add, delete, replace — to a document called best_skill.md, keeping only the versions that actually improve performance on a held-out set. What's interesting isn't the training loop itself. It's what happens when you take that finished file and hand it to a completely different setup.

The standout number comes from a cross-harness test. A skill optimized inside the Codex environment was dropped into Claude Code with zero further tuning, on the SpreadsheetBench benchmark. It pushed Claude Code from a baseline of 22.1 up to 81.8 — edging past the 80.4 that Claude Code managed when it trained its own skill from scratch. Two harnesses with different file APIs and command surfaces, and the borrowed instructions did slightly better than the homegrown ones. The researchers argue this only works because the skill isn't encoding tool-specific command tricks; it's teaching a structural habit — inspect the workbook, verify formulas, write static values across the target range instead of trusting Excel's recalculation.

Math told a different story. Transferring a LiveMath skill from Codex to Claude Code retained just 10% of the in-domain gain, and the reverse direction wasn't much better at 30%. Reasoning-heavy tasks seem to stay glued to whatever environment trained them, while procedural tasks — the kind with a clear sequence of checks and formats — travel well. Cross-model transfer within the GPT-5.4 family showed a similar split: SpreadsheetBench skills kept 82% of their value moving from the full model down to the mini variant, but only 16% surviving the drop to nano.

None of the transferred skills fell below the no-skill baseline, across all eleven transfer rows Microsoft tested. That's a modest but real floor. Training itself isn't free — the paper reports 0.6 million to 46.4 million tokens spent per benchmark point depending on task difficulty — but it's a one-time offline cost. Once the file exists, deploying it anywhere costs nothing extra at inference time. The finished artifacts are also tiny, running 379 to 1,995 tokens, built from as few as one accepted edit.

The scope here is narrower than the headline suggests. Every experiment stays inside the GPT model family; nobody has tried moving a skill from GPT to Qwen or Llama. And the benchmark set is small — six tasks, two per comparison axis. What Microsoft has actually shown is a mechanism and a promising first data point, not a general law of skill portability.