When: the plan is not large enough to run the best model on everything, or the job is big enough that spending the best model on rote work is a waste.
Built from: which model to use, agents, the context window.
Split the job by how much judgement each part needs.
- Plan with the best model you have. Fable on medium. Have it work out what needs doing, in what order, and what "done" looks like. Save the plan to a file.
- Execute with a cheaper model. Hand the plan to an agent on a model that fits the work. Coding can go to Sonnet. Anything that still needs judgement stays on Opus or higher. The agent works from the file, not from your window.
- Review with the best model again. Fable on medium reads the result against the plan. What was done, what was missed, what was done wrong.
The best model touches the job twice, at the two points where judgement matters. The middle, where most of the tokens go, runs on something cheaper.
As a skill: ask the AI to turn this page into a skill that takes a job description and runs all three steps.