Skills before training

Operant·

When a smaller model fails at a recurring task, the reflex is to reach for training. The reflex is usually premature. Most of what separates the frontier model's performance from the smaller model's on a specific, recurring task is not raw capability — it is unstated context: which tools to use and in what order, what the output must look like, which edge cases matter, what to do when a step fails. The frontier model infers those from thin air. The smaller model needs them written down.

A skill is the write-down

That write-down is what Operant calls a skill: a portable markdown artifact, learned from the traces of successful runs, stating the tools and the judgment one task family takes. It is not a prompt hack and not fine-tuning — it is an explicit, inspectable statement of method. You can read it, edit it, export it, and version it. Its provenance records which runs taught it and when.

In our published run, the learned skill is provably not a copy of its seeds: the two seed files hold only method and format, 6,155 bytes together, while the output is 7,270 bytes — more than all the seed material combined, so it cannot be a concatenation of it. And the skill-plus-smaller-model pair scored 0.9667 on held-out exemplars against a 0.8 parity threshold. No weights changed at any point.

Why the ordering matters

Doing skills first is not just cheaper — it is diagnostic. If an explicit skill closes the gap, the gap was context, and training would have spent weeks encoding into weights what a page of markdown states directly. If the skill does not close the gap, you have learned exactly where capability genuinely falls short, and any later training has a precise target and a ready-made acceptance criterion: the same held-out gate the skill just failed.

There is also a control argument. A skill can be inspected before it serves and diffed when it changes. Weights cannot. For a production agent where a silent regression costs real money, start with the intervention you can read — and let the gate, not the roadmap, decide whether the heavier machinery is ever needed.