Microsoft SkillOpt Enables Agent Skill Portability Across Models and Codex and Claude Code
New research from Microsoft reveals that SkillOpt trains reusable natural-language skill documents that transfer seamlessly across different AI models and harnesses.
Stock photo for illustration only, not from the actual event
- SkillOpt trains a single natural-language skill document while keeping the target model frozen.
- Tests confirm that skills optimized in Codex successfully transfer to Claude Code, matching or exceeding native training.
- Exported artifacts consist of compact text files ranging from 379 to 1,995 tokens.
- Training costs are paid once offline, adding zero additional inference-time optimizer calls during deployment.
SkillOpt operates by training a single natural-language skill document while the target model remains completely frozen without weight modifications. An optimizer model reviews scored execution rollouts, proposing bounded additions, deletions, or replacements. Edits are only accepted when validation scores show strict improvement on a held-out split, ultimately exporting a single best_skill.md file.
Evaluation tables across the GPT model family highlight compelling retention metrics. For instance, SpreadsheetBench on the GPT-5.4-mini variant retains 82% of its in-domain performance gain. Meanwhile, the LiveMath benchmark on the GPT-5.4-nano variant yields a surprising score of 28.8, outperforming the in-domain SkillOpt result of 27.2, which the research paper interprets as evidence of model-agnostic procedural learning.
The most crucial finding for practical deployment involves cross-harness transfer between Codex and Claude Code, which utilize vastly different APIs and command structures. Tests demonstrated that a skill optimized inside Codex elevated Claude Code from a baseline score of 22.1 to 81.8, slightly surpassing the 80.4 score Claude Code achieved by training its own skill entirely from scratch.
The successful portability across distinct execution harnesses implies that effective agent skills do not rely on memorizing specific CLI command syntax. Instead, they encode robust workbook-level procedures—such as structure-first inspection, formula-aware verification, and static-value materialization—which remain functional regardless of which environment executes the Python code.
Stock photo for illustration only, not from the actual event
Examining the architectural footprint, the final exported skill artifacts remain remarkably lightweight, ranging between 379 and 1,995 tokens across six benchmarks, with a median close to 920 tokens. These are assembled from just 1 to 4 accepted optimizer edits. For SpreadsheetBench, the verbatim rule focuses on inspecting workbook structures and formulas before writing evaluated static values across the requested range rather than depending on spreadsheet recalculation engines.
Training expenses are incurred precisely once in an offline phase, ranging from 0.6M to 46.4M training tokens per absolute test point depending on the benchmark. SpreadsheetBench requires 0.6M tokens per point, whereas DocVQA requires 46.4M. The optimizer model operates strictly during training and introduces zero inference calls during production deployment.
"A skill optimized inside Codex lifted Claude Code from 22.1 to 81.8. That slightly exceeds the 80.4 Claude Code reached by training its own skill from scratch."
MarkTechPost Research Report
Furthermore, the auditing perspective offers a distinct operational advantage. Because the deployed artifact is a plain text document, domain practitioners can read and review it within minutes. Every modification is fully traceable through recorded edit_apply_report.json files containing per-edit accept and skip statuses, providing a transparent operational posture entirely distinct from shipping opaque fine-tuned model weights.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment