Microsoft researchers, alongside teams from Shanghai Jiao Tong University, Tongji University, and Fudan University, developed SkillOpt, a text-space optimizer that trains a single natural-language skill document while keeping the target model frozen. The optimizer proposes bounded edits, accepting changes only when scores strictly improve, and exports a single file, best_skill.md. Cross-model transfer tests within the GPT-5.4 family showed mixed retention of in-domain gains, with SpreadsheetBench on GPT-5.4-mini retaining 82% of the gain, while GPT-5.4-nano on SpreadsheetBench retained only 16%. Notably, LiveMath on GPT-5.4-nano saw a transferred skill score of 28.8, exceeding its in-domain SkillOpt result of 27.2, suggesting some procedures are target-model agnostic. No row fell below the target’s no-skill baseline, but cross-family transfer was not tested. Cross-harness transfer using GPT-5.5 yielded the strongest result: a skill optimized in Codex lifted Claude Code’s SpreadsheetBench score from 22.1 to 81.8, slightly exceeding the 80.4 achieved by training directly on Claude Code, retaining 102% of the in-domain gain. However, LiveMath portability was low, with Codex → Claude Code retaining only 10% of the gain. The paper attributes SpreadsheetBench’s portability to workbook-level procedures like