SkillOpt creates portable agent skill artifacts that transfer across model scales and harnesses—preserving up to 102% of in-domain gains from Codex to Claude Code.
Building an AI agent that can tackle real-world tasks usually means fine-tuning a model or assembling a stack of custom tools. But what if the secret to a more capable agent were simply a better markdown file? Microsoft’s SkillOpt shows that it can be. The framework optimizes a single natural-language skill document—best_skill.md—while keeping the target model frozen. The payoff is a collection of optimized agent skill artifacts that can move across model scales and between Codex and Claude Code harnesses. The results come from a research paper posted on arXiv (2605.23904v2) and highlighted on MarkTechPost.
SkillOpt is part of a rising wave of text-based agent skill optimization. Rather than updating weights, the system uses an optimizer model to propose edits to an existing skill document. Those edits are bounded: add, delete, or replace. Once the optimizer finds a candidate, it evaluates the modified skill against a target benchmark and keeps improvements.
Because the target model stays frozen, no fine-tuning pipeline is needed. The optimization cost is paid once, offline. At inference time, the model simply reads the optimized skill and follows it—no extra optimization calls, no dynamic search.
Equally important, SkillOpt doesn’t require a bespoke skill format. The shared best_skill.md contract works in direct chat, Codex, and Claude Code. That shared layer is what makes portability possible at all.
Cross-harness transfer is where SkillOpt surprises. It evaluated skills trained inside two popular agent harnesses: Codex and Claude Code. Instead of forcing teams to re-optimize in each environment, the framework produces skills that travel.
Spreadsheet procedures appear to transfer well because they are largely procedural: lookup, format, compute, repeat. A well-crafted set of instructions is useful no matter which harness parses it. All 4 cross-harness transfer rows in the paper land above the target model’s no-skill baseline, so even weak transfer is better than no skill.
Within the GPT-5.4 family, results are more nuanced. On SpreadsheetBench, a skill optimized with GPT-5.4 transferred to GPT-5.4-mini with 82% of its in-domain gain retained. That is strong but not perfect. The same skill transferred to GPT-5.4-nano retained only 16%, a reminder that smaller models may not be able to follow the full procedure encoded in the artifact.
Yet the LiveMath results break the simple story. A GPT-5.4-trained LiveMath skill applied to GPT-5.4-nano scored 28.8, higher than the in-domain optimized score of 27.2—roughly 140% retention. Some learned procedures appear to be target-model agnostic; in this case, the transferred skill provided a surprisingly good inductive bias for a much smaller model.
The key insight is that transfer strength isn’t just a function of model size. Task type and skill structure matter as much as the model family. Every cross-model row in the paper still beat the no-skill baseline.
Combine the cross-harness and cross-model results, and a clear pattern emerges: procedural tasks transfer strongly; reasoning-heavy tasks transfer weakly. SpreadsheetBench skills are instruction-friendly: they tell the agent to read cells, apply formulas, and handle edge cases. Math skills depend on multistep reasoning, which is more sensitive to the exact model’s strengths.
The paper’s cross-benchmark results add another layer. Transferring an OlympiadBench skill to Omni-MATH improved GPT-5.4 from 56.6 to 60.3, a gain of +3.7 points. That suggests relevant skills can lift performance on adjacent benchmarks, not just the exact task they were written for.
Across all transfer combinations—4 cross-model, 4 cross-harness, and 3 cross-benchmark—every row landed above the target model’s no-skill baseline. Even the weakest transfer added value; stronger transfer approached parity or exceeded it.
SkillOpt also solves a practical deployment issue: portability is worthless if artifacts are unwieldy. The final skills in the paper ran from 379 to 1,995 tokens, with a median near 920. That is small enough to version in Git, review in a pull request, and share across teams.
Each optimized skill was assembled from just 1 to 4 accepted edits. The compact footprint is a direct result of bounded edit proposals and a clean natural-language contract.
Training cost is also contained. SkillOpt requires a one-time offline optimization pass using 0.6M to 46.4M training tokens per absolute test point, depending on the benchmark. SpreadsheetBench sits at the low end at 0.6M; DocVQA sits at the high end at 46.4M. At inference time, there are zero optimization calls—the deployed model simply follows the skill.
SkillOpt’s approach fits a growing trend in agent development: move complexity from model weights to structured prompts and skill files. This has several practical consequences.
best_skill.md may emerge as a de facto interchange format, the way prompt files have become common in agent repositories.A skill marketplace is the natural next step. Teams could publish best_skill.md files the way they publish open-source prompts today.
The paper is not a blanket proof that all agent skills are instantly portable. Its evidence is limited to one GPT family and two benchmarks per transfer axis. Cross-harness tests cover Codex and Claude Code only. The authors have not yet demonstrated broad generalization across open-weight models, other task families, or more diverse agent harnesses.
Reasoning-heavy skills, in particular, need more work. The 10% retention result for Codex-trained LiveMath in Claude Code is a clear warning that transfer depends on the task. Optimizing for portability may require explicit regularization or ensemble-style methods.
Still, the direction is promising. Text-based skill optimization is rising across the 2026 research cycle, and cross-harness portability has now been demonstrated in a public preprint.
Microsoft’s SkillOpt shows that optimized agent skill artifacts can transfer across model scales and between Codex and Claude Code harnesses. The headline number—102% retention from Codex to Claude Code on SpreadsheetBench—is an eye-catching proof point. But the deeper lesson is that agent capability can live in a small natural-language file, not just in model weights.
Teams building AI agents should watch this space. If the pattern holds across more models and benchmarks, SkillOpt-style optimization could make agent skills portable, inspectable, and affordable. For now, the practical takeaway is simple: when optimizing agents, don’t ignore the artifacts. A well-written best_skill.md may be the most transferable asset you have.
SkillOpt is a framework that optimizes a natural-language skill document—typically called best_skill.md—instead of updating the model's weights. It uses an optimizer model to propose bounded edits like additions, deletions, or replacements, then evaluates the modified skill against a benchmark to keep improvements. Because the target model stays frozen, SkillOpt avoids the cost and complexity of a fine-tuning pipeline.
SkillOpt creates skills in a shared best_skill.md format that does not depend on a single harness, so the same artifact can be read and followed by different agent harnesses at inference time. In their tests, a Codex-trained SpreadsheetBench skill scored 81.8 inside Claude Code, beating Claude Code's own 80.4 in-domain score. This 102% retention shows that a skill optimized in one harness can be dropped into another without re-optimizing.
No. SkillOpt keeps the target model frozen and only changes the skill document that guides the agent. The optimization happens once offline, and at inference time the model simply reads the optimized skill and follows it. This means you can apply SkillOpt to an existing model without additional fine-tuning or dynamic search at runtime.
The paper found that spreadsheet-related procedures, such as the SpreadsheetBench skill, transfer very well, with up to 102% retention from Codex to Claude Code. In contrast, reasoning-heavy tasks like LiveMath showed weaker transfer, dropping to 30% retention from Claude Code to Codex and only 10% from Codex to Claude Code. This suggests that procedural, rule-based skills are more portable, while complex reasoning skills are more sensitive to the model or harness that executes them.
The main benefit is that you can optimize a skill once in one harness and deploy it across multiple models and agent environments without repeating the expensive optimization process. This reduces engineering effort and makes it easier to share reusable agent capabilities across teams and products. It also opens the door to a future where agents improve by swapping better skill artifacts rather than replacing or retraining the underlying model.
