## Article: SkillOpt Transferability—How Much of a Learned Skill Survives the Move?
**SkillOpt** is a text-space optimizer created by a team of researchers from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University. Instead of fine-tuning the target model, SkillOpt trains a single, human-readable skill document while the target model remains frozen. An optimizer model proposes bounded add/delete/replace edits based on scored rollouts, and a held-out selection split accepts an edit only when the score strictly improves. The output is a single file, `best_skill.md`.
—
### What the Transfer Tables Show
The reported transfer tables include three columns:
– **Baseline**: The target’s no-skill score
– **Direct**: SkillOpt trained in-domain on that exact target
– **Transferred**: A skill trained elsewhere, applied without further optimization
The key insight is that the most useful comparison is not *transferred vs direct*. Instead, the paper asks: **how much of the in-domain gain survives when the skill is moved to a new environment?**
—
### Cross-Model Transfer Within the Same Family
Skills were trained on GPT-5.4 and deployed on smaller variants without any re-optimization.
| Benchmark | Target | Baseline | Direct | Transferred | Gain | Retained |
|——————-|————–|———-|———|————-|——–|———-|
| SpreadsheetBench | GPT-5.4-mini | 36.1 | 47.5 | 45.5 | +9.4 | 82% |
| SpreadsheetBench | GPT-5.4-nano | 23.5 | 42.5 | 26.5 | +3.0 | 16% |
| LiveMath | GPT-5.4-mini | 14.7 | 32.8 | 19.2 | +4.5 | 25% |
| LiveMath | GPT-5.4-nano | 23.2 | 27.2 | 28.8 | +5.6 | 140% |
**Takeaway:**
– Spreadsheet skills transfer extremely well, retaining up to **82%** of in-domain gains.
– Math-heavy skills transfer less reliably.
– All transferred skills remain above the no-skill baseline.
– Scope is limited to one model family; cross-family transfer is not tested.
—
### Cross-Harness Transfer (Strongest Real-World Result)
All rows use GPT-5.5, with different harnesses exposing different tool and file APIs. A skill that survives this shift is not encoding command recipes—it encodes general procedure.
| Benchmark | Source → Target | Baseline | Direct | Transferred | Gain | Share of In-Domain Gain |
|——————-|———————-|———-|——–|————-|——-|————————-|
| SpreadsheetBench | Codex → Claude Code | 22.1 | 80.4 | 81.8 | +59.7 | 102% |
| SpreadsheetBench | Claude Code → Codex | 27.5 | 85.0 | 71.1 | +43.6 | 76% |
| LiveMath | Claude Code → Codex | 35.2 | 78.4 | 48.0 | +12.8 | 30% |
| LiveMath | Codex → Claude Code | 40.8 | 56.5 | 42.4 | +1.6 | 10% |
**Takeaway:**
– The Codex → Claude Code SpreadsheetBench result is the headline: transferred performance slightly exceeds in-domain training.
– Procedural spreadsheet skills travel well; reasoning-heavy skills do not.
– The artifact’s shared format (`best_skill.md`) makes cross-harness portability possible.
—
### Cross-Benchmark Transfer: Real but Small
With no in-domain SkillOpt run reported, only no-skill baselines are available.
| Source → Target | Model | Baseline | Transferred | Gain |
|———————–|———|———-|————-|——-|
| OlympiadBench → Omni-MATH | GPT-5.4 | 56.6 | 60.3 | +3.7 |
| OlympiadBench → Omni-MATH | GPT-5.4-mini | 34.8 | 36.6 | +1.8 |
| OlympiadBench → Omni-MATH | GPT-5.4-nano | 38.8 | 40.1 | +1.3 |
**Takeaway:**
– Small but consistent gains.
– Interpreted as retained mathematical procedure despite format changes.
—
### Why the Artifact Moves at All
All execution modes consume the same `best_skill.md` format. The Codex harness produces a per-task `SKILL.md` and compact execution trace; the Claude Code harness mirrors the workspace contract through the `claude` CLI. Neither harness receives a bespoke skill format.
Final skills range from 379 to 1,995 tokens (median ~920), assembled from 1 to 4 accepted edits. The paper highlights procedural rules—e.g., “inspect workbook structure and formulas, then write evaluated static values”—as the portable unit.
—
### What This Implies for Portability
– **Training cost is paid once** offline and measured in tokens, not deployment calls.
– Skills trained in one harness can be deployed in another, enabling optimization where tooling is cheapest.
– The artifact is inspectable—a text file readable by domain practitioners—with traceable edits.
—
### Key Takeaways
– Portabilidade está demostrada, pero aún no generalizada: solo una familia de modelos y dos benchmarks por eje.
– Una habilidad de hoja de cálculo entrenada con Codex obtuvo 81,8 en Claude Code, superando el resultado en-domain de ese entorno (80,4).
– Todos los 12 resultados de transferencia (4 cruz-modelo, 4 cruz-armazon, 3 cruz-benchmark) permanecen por encima del baseline sin habilidad.
– La fuerza de la transferencia sigue el tipo de tarea: habilidades procedimentales (hojas de cálculo) se mueven bien; habilidades de razonamiento, menos.
– La unidad portátil es un solo `best_skill.md` de 379 a 1,995 tokens, construido con 1 a 4 ediciones aceptadas.
—
### FAQ
**What is SkillOpt?**
SkillOpt is a text-space optimizer that trains a natural-language skill document while keeping the target model frozen. It iteratively proposes edits and accepts only those that strictly improve held-out scores. The output is a single file, `best_skill.md`.
**How does SkillOpt differ from fine-tuning?**
Unlike fine-tuning, SkillOpt does not update model weights. Instead, it produces a skill document that the model consumes at inference time, enabling reuse across architectures without retraining.
**What counts as a “transfer” in the paper?**
A transferred skill is trained in one environment (model, harness, or benchmark) and evaluated in another without further optimization. The baseline is the target’s no-skill score.
**Why is “Direct” included if the skill is transferred?**
“Direct” represents the best possible in-domain performance with SkillOpt. Comparing “Transferred” against “Direct” shows how much performance is lost when moving skills across contexts.
**What explains strong cross-harness results?**
Cross-harness success suggests the skill encodes procedural, workspace-level behavior rather than command-specific syntax. The shared `best_skill.md` contract makes this possible.
**Are cross-family transfers tested?**
No. The study limits itself to within-family transfers (e.g., GPT to GPT variants). Cross-family portability remains an open question.
**Is portability always high?**
No. Spreadsheet skills transfer strongly, while math-reasoning skills transfer weakly. Task type and procedural regularity strongly influence retention.
**What makes a skill portable?**
Portable skills are procedural, inspection-friendly, and model-agnostic. They rely on stable workflows (e.g., inspect → verify → materialize) rather than environment-specific reasoning.
**How efficient is SkillOpt?**
Training cost is paid once, with token counts ranging from 0.6M to 46.4M per benchmark. At deployment, the optimizer adds zero runtime overhead.
**Where can I read more?**
The paper (arXiv 2605.23904v2), GitHub project, project page, documentation, and demo video are available as primary resources.
—
### Conclusion
SkillOpt demonstrates that large-language-model skills can be trained once and reused across models, harnesses, and benchmarks—when the task is procedural and the artifact is text-based. The strongest result shows a Codex-trained spreadsheet skill outperforming in-domain training inside Claude Code, validating a practical path toward efficient, inspectable, and portable AI skill deployment. While not universally applicable, SkillOpt offers a compelling alternative to repeated fine-tuning, especially in multi-environment or constrained-compute settings.
