Use expensive models to discover structure, then cheap models/code to run it repeatedly. Let the big model figure out categories, rubric, workflow, edge cases, and architecture.
Once the problem stabilizes, move repetitive execution to cheaper inference or deterministic code.