A text-space optimizer trains agent skill documents like model weights

Most agent skill files are hand-written, generated once, or lightly self-revised. SkillOpt instead treats the skill document as a trainable artifact: a frozen target model runs rollouts under the current skill, a separate optimizer model proposes bounded add, delete, and replace edits, and an edit is kept only when it improves a held-out validation score. The authors report the method was best or tied across all 52 combinations they tested, spanning six benchmarks, seven target models, and three harnesses (direct chat, Codex, and Claude Code), beating four existing skill-training approaches (Trace2Skill, TextGrad, GEPA, EvoSkill). On GPT-5.5, they report it raised accuracy over no skill by 23.5 points in direct chat, 24.8 points inside Codex, and 19.1 points inside Claude Code, and that a skill trained inside Codex kept most of its gain when moved into Claude Code without further training. These are the authors' own benchmark numbers from a preprint; this run did not rerun the code to verify them.

Why it matters: For anyone maintaining an AGENTS.md-style skill file, the result argues that a skill document behaves like a tunable parameter rather than fixed prose, and that gains made training under one harness can carry into another without redoing the work.