Simon Willison, the developer behind Datasette and a prominent voice on practical AI engineering, argued in a July 3 post on his blog, simonwillison.net, that builders should describe less and delegate more when instructing Fable, the planning-oriented mode inside Claude Code. The claim traces back to a fireside chat he hosted with Cat Wu and Thariq Shihipar of the Claude Code team at the AI Engineer conference. Their example was testing, and it reframes a habit many teams have never questioned.
Instead of instructing Fable to run automated tests only on larger features while skipping them for minor cosmetic edits, the Claude Code team’s advice is to hand that decision to Fable itself. Willison passes along the shift plainly: drop the conditional rule, and simply ask Fable to gauge, on a case-by-case basis, whether a change warrants a test.
A second, related tip in the post extends the same logic to model selection. Jesse Vincent, a builder Willison credits with the suggestion, proposed routing smaller coding tasks to cheaper models, with Fable itself picking which model fits each task. Willison frames the timing as partly financial: Fable’s token allowance was on track to shrink faster once prices rose in the days following the post.
Willison tested the idea on himself. He told Claude Code, in a single line, to use its own judgment to pick an appropriately less powerful model and hand off each coding task to a subagent running it. Claude Code responded by writing a memory file to its project folder, capturing the new policy so it would persist across future sessions without Willison repeating the instruction.
That memory file, not the one-line prompt, is the interesting artifact. Its contents assign implementation work to a mid-tier model and mechanical edits to the cheapest one, while reserving design and review judgment for the top-tier model already running the session. Willison did not write that breakdown. Claude Code derived it and saved it as its own standing policy.
This is the reverse of how most teams write agent instructions today, which tend toward exhaustive conditionals: test if the diff exceeds some line count, skip if it only touches copy, escalate to a bigger model if the task looks ambiguous. Willison’s post argues that this enumeration is often wasted effort, because the model already holds the context needed to make the call, and forcing it to also memorize a rulebook adds overhead without adding accuracy.
The operational lesson is narrow enough to test this week. Builders maintaining long conditional blocks in their Claude Code instructions should try deleting one rule, replacing it with a single sentence granting judgment, and watching whether output quality holds over the next few sessions. If it does, that is fewer prompt-tuning hours per project and fewer premium-model tokens spent on routine work before the price increase Willison flagged takes effect.
Simon Willison detailed this approach in a July 3, 2026 post on simonwillison.net.
