Judgment Over Dictation in Claude Code

Main idea

The Claude Code team shared a tip that Willison considers a fundamental shift in approach: instead of dictating steps, let the model decide what is appropriate — including delegating smaller coding tasks to cheaper subagent models.

Context

Willison shared this on July 3, 2026 — as many developers experiment with agentic workflows and optimize performance/cost ratios. Example prompt: For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent.

Why it matters

For developers using Claude Code, this is a concrete, immediately actionable tip to reduce token costs without losing quality. It illustrates a broader shift: trusting AI for operational decisions while humans retain control over high-level direction.

Details and arguments

• Instead of prescriptive test rules, let Fable decide when testing is appropriate
• Delegating coding tasks to Sonnet and Haiku preserves expensive Fable tokens
• Fable tokens reserved for judgment-heavy work: design, audit, synthesis
• Result: faster productivity and lower costs