Microsoft’s SkillOpt enables AI skill transfer across models and tools
A single text file may soon let AI agents carry learned skills across different models and tools without retraining. Microsoft’s SkillOpt, developed with Shanghai Jiao Tong University, Tongji University, and Fudan University, trains a natural-language skill document that can be exported as best_skill.md and reused across model scales and even between coding environments like Codex and Claude Code.
How portable skills work under the hood
SkillOpt freezes the target model and uses an optimizer to edit a skill document through bounded add, delete, and replace operations. Each edit is accepted only if it strictly improves a held-out score. The result is a compact artifact that can be transferred and evaluated immediately. The research highlights two key dimensions of portability: within-family (same model series) and cross-family (different models or environments).
Cross-model and cross-harness results show uneven gains
Within the same GPT family, transfer rates vary widely. On SpreadsheetBench, a GPT-5.4-mini model retained 82% of the in-domain gain after receiving a skill trained on GPT-5.4, while the smaller nano variant kept only 16%. In contrast, LiveMath on GPT-5.4-nano showed a surprising 140% gain over the in-domain baseline, suggesting some procedural skills are model-agnostic. The strongest evidence, however, comes from cross-harness transfer. A skill trained in Codex and transferred to Claude Code lifted performance from 22.1 to 81.8 on SpreadsheetBench—slightly surpassing the 80.4 achieved by training directly in Claude Code. Yet, the same transfer yielded only 10% retention for LiveMath, indicating that reasoning-heavy tasks remain tightly coupled to their training environment.
Why it matters
SkillOpt’s portability claims are uneven but meaningful: procedural skills like structure inspection and formula verification transfer more reliably than reasoning tasks. For teams deploying AI agents across models or tools, this could reduce redundant training cycles and speed up deployment. The research stops short of claiming universal transferability, but the ability to move skills between Codex and Claude Code—two distinct coding harnesses—hints at a future where skill artifacts become standardized, portable interfaces for AI agents. The uneven retention rates also serve as a reminder: not all skills are created equal, and the stickiest ones may be the simplest to move.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.