Developer documents workflow to let coding agents exercise their own judgment

Simon Willison reports adopting a prompt strategy that instructs the coding agent Fable to exercise its own judgment when deciding whether to write tests and which subagent model to use for a given task.

He cites a tip from the Claude Code team suggesting it is more effective to tell Fable to "use its own judgement" about testing than to hard-code rules like "only use automated testing for larger features."

Willison also describes a related tip from Jesse Vincent to reduce token burn by instructing Fable to apply its own judgment about which model to use for smaller tasks.

He implemented the guidance by prompting Claude Code with: "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent."

Claude Code saved the instruction as a memory file at ~/.claude/projects/name-of-project/memory/delegate-coding-to-subagents.md, specifying that judgment-heavy work such as design, auditing, data synthesis, and review should remain in the main model while implementation edits can be delegated to subagents.

The memory file names Sonnet for substantive implementation and Haiku for trivial or mechanical edits, with results reviewed in the main loop before committing.

Willison states the approach is yielding faster progress and slower depletion of his Fable token allowance.