SkillOpt Transfers Agent Skills Between Codex, Claude

A team from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University has published a method called SkillOpt that treats agent instructions as a single tunable document rather than a fixed prompt. The target model itself never gets retrained. Instead, an optimizer model reads scored rollouts and proposes bounded edits, additions, deletions, replacements, to a natural-language skill file.

A held-out split decides whether each edit sticks: it only survives if the score strictly improves. The end product is one exported file, best_skill.md, meant to travel.

The interesting question isn't whether SkillOpt beats a no-skill baseline in-domain. It's what happens when that trained skill file gets moved somewhere else entirely, a smaller model, a different harness, without any further optimization. The paper's transfer tables track exactly this: baseline score, direct in-domain score, and transferred score, then ask how much of the direct gain survives the jump.

Cross-model results within the GPT-5.4 family already show that retention is uneven, ranging from near-total on SpreadsheetBench to weak on GPT-5.4-nano. But the harder test, and the one the researchers treat as their strongest evidence, is what happens when the skill file moves not just across model sizes but across agent harnesses entirely.