Let Claude pick a cheaper model for the boring parts

Most advice for steering coding agents is about writing tighter instructions. Simon Willison has landed somewhere different: with a capable model like Claude Fable, he gets better results by handing over judgement rather than rules. Instead of a brittle instruction such as "only use automated testing for larger features, don't run tests for small copy or design changes," he now just tells the model to "use its own judgement when deciding to write tests." The model reads the situation better than a rigid rule can describe it.

The more interesting move is what this does to cost. Willison adds a single line to his setup: "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent." Fable then keeps the work that actually needs its strength, design, auditing, and pulling results together, and farms the rest out. Substantial implementation goes to Sonnet, trivial edits go to Haiku, each running as a subagent whose output Fable reviews in the main loop before anything gets committed. He reports getting a lot done while his Fable token allowance drains more slowly than it used to.

Why it matters

If you pay for a premium model and burn through its allowance on routine edits, this is a small change with a direct payoff: one prompt that lets the expensive model delegate the cheap work while still checking it. The tradeoff is trust, since you are letting the model decide when a weaker one is good enough, so keep the review-before-commit step rather than skipping it.