Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses
MarkTechPost Asif Razzaq ● Covered by 5 sources
Microsoft researchers built SkillOpt, a system that trains one text file of agent instructions instead of retraining a model. That file can jump between different AI models and even different coding tools and still work.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
SkillOpt does something deceptively simple: it optimizes a single natural-language document, best_skill.md, while leaving the underlying model completely untouched. An optimizer model looks at scored attempts, proposes edits, and a held-out test only accepts a change if it strictly improves the score. What comes out the other end is one file, built from as few as one and as many as four accepted edits, running anywhere from 379 to 1,995 tokens depending on the benchmark.
The interesting part isn't that this works in-domain. It's what happens when you move the trained skill somewhere else entirely. Researchers from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University tested transfer across three axes: smaller model variants, different agent harnesses, and different benchmarks. On SpreadsheetBench, a skill trained on GPT-5.4 and moved to the smaller GPT-5.4-mini kept 82% of the in-domain performance gain. On GPT-5.4-nano, the same benchmark only held onto 16%, so retention swings a lot depending on the target.
The headline result, though, is cross-harness. A skill optimized entirely inside the Codex environment was dropped into Claude Code with zero further tuning, and SpreadsheetBench performance jumped from a no-skill baseline of 22.1 to 81.8. That's slightly better than the 80.4 Claude Code managed when it trained its own skill from scratch. Codex and Claude Code expose different tool APIs and different command surfaces, so a skill surviving that jump is not just memorizing command syntax — it's carrying over something like a workbook-level procedure: inspect structure first, verify formulas, and write out static values rather than trusting spreadsheet recalculation.
Math reasoning didn't transfer nearly as well. Going from Codex to Claude Code on the LiveMath benchmark preserved only 10% of the in-domain gain, and the reverse direction wasn't much better at 30%. The pattern that emerges is that procedural know-how — how to check, verify, and format — travels across environments reasonably well, while reasoning-heavy skills stay more tightly bound to wherever they were trained. Cross-benchmark transfer, tested by moving a skill from OlympiadBench to Omni-MATH, produced consistently positive but modest gains across three model sizes, without ever collapsing below baseline.
All of this rests on a shared file format: Codex and Claude Code both consume the same best_skill.md, rendered into a per-task SKILL.md and read back through a compact execution trace. Training happens once, offline — the paper reports training costs ranging from 0.6M tokens per test point on SpreadsheetBench up to 46.4M on DocVQA — and the optimizer adds nothing extra at deployment time. Every edit gets logged in an accept/skip report, so unlike shipping new model weights, what changes and why stays auditable in a document a person can actually read.
My take — AI-written commentary, not fact-checked reporting
The auditability angle is the part people will undersell in favor of the flashy transfer numbers, but it's arguably the more important story. A readable text file with a logged edit history beats a black-box fine-tune for anyone who has to explain to a compliance team why the agent suddenly behaves differently. That said, the evidence here covers one model family and a handful of benchmarks, so treat the portability claim as promising, not proven — especially since the math-reasoning results show this approach clearly has limits.
Read more about this at: MarkTechPost