When output quality disappoints, the instinct is to rewrite the prompt. Frequently the prompt is fine and the model is wrong for the task. Long-document reasoning, fast classification, and image work have genuinely different best answers, and no amount of wording closes that gap.
Routing also changes the economics. A pipeline that sends every step to the largest available model is paying premium rates for work a smaller one handles identically. Classification, extraction, and formatting steps rarely need frontier capability. Reserve that for the steps where judgement actually matters.
The practical approach is to decompose the workflow into steps, identify what each step genuinely requires, and assign the cheapest model that clears the bar with margin. Then measure. The margin matters because model updates shift capability in both directions, and a step running at the edge of what a small model can do will fail first.
This is unglamorous work compared to prompt craft, but it is where most of the available improvement sits for teams already past the basics.