Stop Writing Rule Lists for Your Coding Agent
Anthropic's Claude Code team told Simon Willison: stop dictating cases, and let the agent pick which cheaper model handles the work.
Simon Willison has a CLAUDE.md file he can paste into any project that tells his coding agent, Anthropic's Claude Code running on its Fable model, to use its own judgment, including judgment about which cheaper sibling model should handle the work. That single instruction, persisted by Claude Code itself as a project memory file, is the cleanest artifact yet of where Anthropic's coding tool is pushing practitioners: away from rule-lists, toward agent-controlled judgment.
The advice came out of a Fireside Chat Willison hosted at the AI Engineer (AIE) conference on July 1, 2026, with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team. Willison published the takeaways on July 3. Their central recommendation: stop writing exhaustive case-by-case rules, and instead tell the model to use its own judgment.
The example they gave was testing. The instinct is to instruct the model: "only use automated testing for larger features, don't update and run tests for small copy or design changes." The Anthropic team's preferred framing is to tell the agent to use its own judgment about when to write tests. The mental model is delegation with review, not abandonment.
Developer Jesse Vincent, in a tip relayed by Willison, pushed the idea one layer down: instruct the agent to delegate smaller coding tasks to subagents (secondary AI workers that the main agent can dispatch a task to) running lower-power models, letting the main model pick which sibling to invoke. Willison tried a prompt that read, in effect, "use your judgement to decide an appropriate lower power model and run that in a subagent." Claude Code saved the rule as a project memory file named delegate-coding-to-subagents.md. The artifact is the proof: the convention became project-level policy without the user writing one.
The discipline is not free. It lands as secondary outlets flag a Claude Fable pricing and usage-credits change effective July 7, 2026, with Fable 5 already repositioned as the current Claude Code default on June 30, 2026. Willison's own running note from the post: his "Fable allowance is shrinking less quickly than before." The pattern is partly a coping mechanism for a pricing change, a real one, not a capability breakthrough.
Two caveats earn their place. "Use your judgment" still requires the user to review the subagent's work in the main loop; Willison's own memory file says exactly that. And the Anthropic team's framing reaches readers through an independent practitioner, not as on-the-record quotes. The artifact, the persisted memory file, is the verifiable record. Willison's earlier June 22 walkthrough of porting Moebius to a browser with Claude Code shows the pattern in working practice, not just theory.
The Anthropic Claude Fable product page lays out the model family: Opus, Sonnet, Haiku, in descending capability, that Willison is asking his agent to navigate. Subagent delegation across that hierarchy is what "let the agent pick" looks like in practice. Pricing kicks in on July 7. That is the date the discipline gets priced in.