Main idea

Extracting numbers from a list and synthesizing a competitor record from fifteen analyses don't call for the same capability. Yet a skill is labelled by its most demanding step, and everything else it does inherits that tier — two to three times the cost, for an identical result.

The useful carve-up assigns a model per step: structured extraction to a small model, synthesis to a mid-tier model, multi-phase orchestration to the most capable. What justifies the high tier is never the skill as a whole, it is the two or three steps that would fail without it.

Why it matters

This turns a global choice, taken once and never reopened, into a series of local choices you can argue about — and it forces you to say which step justifies the most expensive model.

It is also the lowest-effort lever: it requires no rewriting of the logic.

Nuances and limits

Serving several models within one skill adds plumbing and context handovers between steps. Below a certain volume, the gain doesn't pay for that complexity.

And a step looks simple because it has always been handled by a capable model: downgrading it produces plausible outputs whose error only shows up downstream.

Open questions

• How do you recognize that a step has been downgraded too far, when its result stays plausible?