Skill Artifact Transfer
A skill document optimized against one model, one harness,
and one benchmark turns out to keep working when moved to
another. SkillOpt reports its best_skill.md
artifacts transferring across model scales, between the Codex and Claude
Code harnesses, and to nearby benchmarks — with no further optimization.
If that holds, the optimized skill is not a fitted parameter of the
model it was trained on. It is a statement about the task, and
the training loop was a search procedure for finding it.
That distinction is the whole weight of the concept, because it decides what an optimized skill is worth. A parameter fitted to GPT-5.5 is worth nothing when the model changes next quarter — and models change. A crisp externalization of how to do a task is worth something to every model that reads it, which is why Public Skill Adoption works at all: a repo of two markdown files transfers capability between strangers. Transfer is the property that turns an expensive offline optimization into a durable, distributable asset, and it is the property that makes the offline cost sane. Amortized over one model version, Text-Space Optimization is a curiosity; amortized over every model and harness that will ever read the file, it is an investment.
It also cuts the other way, and the source does not dwell on this: if the artifact transfers, its gains were never model-specific to begin with, which raises the question of how much of the reported lift is optimization discovering something subtle versus optimization discovering the task's obvious structure and writing it down cleanly — something a careful human author might also produce. Transfer is evidence for portability. It is also, quietly, evidence that the thing being learned is generic. The repo's claim of "best or tied-best on all 52 cells" does not settle which.
Claims
- An optimized skill artifact transfers across model scales, across execution harnesses (Codex ↔︎ Claude Code), and to nearby benchmarks without re-optimization. observation — SkillOpt's claim, and the one most worth an independent check: it is the load-bearing empirical premise for treating skills as durable assets rather than fitted parameters.
- What transfers is not fitted to the model, so it is a statement about the task, not about the model. principle — durable, if the observation holds. It is the reason a markdown file can carry capability between agents that share no weights.
- Transfer is what makes offline optimization economic: the cost is paid once and amortized over every model and harness that later reads the artifact. principle — durable. A skill that had to be retrained per model release would cost more than it saves.
- Prefer optimizing the artifact over tuning the model when the target model is frozen, changes often, or is not yours. best practice — context: agents built on hosted frontier models, where weights are inaccessible and deprecation cycles are short. The advice inverts where you own the weights and the task distribution is stable — there, fine-tuning is the cheaper deployment.
- Transfer is double-edged evidence: an artifact that generalizes across models was not exploiting model-specific structure, which weakens the claim that the optimizer found something a good author could not. observation — a tension the source leaves open; the ablations in the paper, not the README, are where it would be resolved.
Related
- Text-Space Optimization — the process that produces the artifact; transfer is what makes the process worth running.
- Execution Commoditization — if models converge and methods transfer, the method file is the asset and the model is a substrate.
- Public Skill Adoption — transfer is the precondition for the whole clone-a-skill economy: someone else's markdown works in your agent because skills are not model-fitted.
- Agentic Distribution — once an artifact transfers, distributing it becomes the problem.
- Pure Agent Application — the artifact is markdown and nothing else, which is exactly why it moves between harnesses.
- Minimal Skill Surface — a 300–2,000 token artifact is small enough to read, audit, and port by hand; smallness and portability reinforce each other.
- Validation-Gated Update — the gate proves improvement on the training harness; transfer is the claim that the improvement was not gate-shaped.
- Golden Templates — the same "an artifact, once proven, is reusable" instinct, arrived at by curation rather than optimization.
- Distillate: microsoft/SkillOpt — training skills like weights, without touching weights