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

SkillOpt is a text-space optimizer that trains a single natural-language skill document while the target model stays frozen. An optimizer model proposes bounded edits to the skill based on scored rollouts, and a held-out split accepts edits only if the score strictly improves. The exported artifact is a single file, best_skill.md. Cross-model transfer within the GPT-5.4 family shows mixed retention: SpreadsheetBench on GPT-5.4-mini retains 82% of the in-domain gain, while GPT-5.4-nano retains only 16% on SpreadsheetBench; LiveMath on nano shows 140% retention, suggesting some procedures are model-agnostic. Cross-harness transfer using GPT-5.5 yields the strongest result: a skill optimized in Codex lifts Claude Code on SpreadsheetBench from 22.1 to 81.8, exceeding Claude Code's own in-domain score of 80.4. However, LiveMath transfer from Codex to Claude Code retains only 10% of the gain, indicating reasoning-heavy skills are less portable. Cross-benchmark transfer from OlympiadBench to Omni-MATH shows small but positive gains across three model scales. The mechanism enabling transfer is that all execution modes consume the same best_skill.md format, and the skill's procedural nature (e.g., workbook structure inspection) makes it portable. Training cost is paid once offline, with no inference-time overhead. The artifact is auditable as a text file with per-edit logs.