Main idea

Converting JSON to CSV, renumbering an index, copying a frontmatter across, sorting a list: these operations have a single result, deducible from their inputs. Handed to the model, they go out through the most expensive channel — generation — and throw in a variability nobody needed. Handed to a script, they cost zero tokens and run in milliseconds.

On competitor_analyze, two steps moved to Python — writing the frontmatter and computing the index delta — removed 111,000 units from consumption, without changing any of the deliverables produced.

The dividing criterion isn't the difficulty of the task but the presence of a judgment: if the result follows from the inputs, there is nothing to have decided.

Why it matters

This gives a rule for carving up a skill's design: every step asks what it expects from the model, and a step that expects no arbitration has no business being addressed to it.

The gain isn't only economic, either: what a script produces is reproducible, what a model produces is only approximately so.

Nuances and limits

Writing and maintaining the script has a cost, invisible in a token report. For a one-off transformation that will never be replayed, going through the model remains cheaper than going through development.

And some transformations look deterministic right up to the first malformed piece of data — the model absorbs the irregularity where the script stops.

Open questions

• Where do you put the boundary when a transformation is ninety percent mechanical and ambiguous for the rest?