Fable's Judgment: Efficiency Tips Simon Willison Gained from the Claude Code Team

Fable now handles more of its own decision making after Simon Willison tested advice from the Claude Code team. The change replaces fixed rules with open prompts that let the model judge when to add tests or call other models.

Willison wrote about the experiments on his blog after a session at AIE. The core shift is simple. Instead of telling Fable exactly when to write tests, he now asks it to decide for itself. The same approach applies to other routine steps.

The second change focuses on cost. Willison routes smaller edits to Sonnet or Haiku while Fable keeps the main loop for judgment and review. Token spend dropped quickly once both patterns were saved in the Claude Code memory file.

These steps address two clear pressures. Model prices are rising. Token budgets on repeated agent runs grow fast. Giving the model room to judge and handing off mechanical work reduces both problems without extra prompting layers.

Letting the model choose its own test timing

Willison replaced detailed checklists with a single request. He now tells Fable to decide when tests add value. In practice this means the model writes tests only when the task involves logic that is easy to break later.

Early runs showed fewer unnecessary test files. The model also caught edge cases that fixed schedules missed. Willison noted that the change kept output quality steady while cutting steps that did not change the final result.

The approach works because current models already carry enough context to judge test value. Removing the rule set lets them use that context directly. Willison stored the updated prompt in the persistent memory file so every new session starts from the same instruction.

Routing small tasks to lighter models

The second adjustment splits work by model size. Fable stays in charge of planning, auditing, and combining results. Sonnet handles the main implementation steps. Haiku takes simple renames, formatting, and other mechanical edits.

Willison reports that this split lowered total tokens per session by a noticeable margin. The main agent still sees every change before it is accepted, so errors from the lighter models are caught early. The pattern is saved alongside the judgment prompt so it runs automatically.

This method matches rising price schedules. As larger models cost more per token, the savings from lighter models compound on any workflow that repeats several times a day.

Why fixed rules often cost more tokens

Detailed instructions increase prompt length. Each added rule consumes tokens on every turn. When the model follows those rules strictly, it performs extra steps even when they are not needed. The extra output then costs more tokens on the return pass.

Open prompts reduce this overhead. The model applies its own judgment once and moves on. Willison found that the net token count fell even though the model sometimes produced longer reasoning traces. The reasoning stayed internal and did not trigger follow-up calls.

The same pattern applies to model choice. Keeping every step on the largest model multiplies cost by the number of turns. Routing only the high-judgment parts to Fable and the rest to smaller models cuts that multiplier.

Early results and remaining limits

Willison has run the updated setup across several weeks of daily agent work. Token usage fell steadily once the memory file held both the judgment prompt and the routing rule. Output quality remained consistent with earlier runs that used more prescriptive instructions.

One limit is still visible. The lighter models occasionally miss context that matters for a small edit. Fable catches most of these cases during review, but the review step itself adds a turn. The net saving stays positive, yet the margin narrows on tasks that require heavy context.

Another limit is prompt drift. If the memory file is cleared or the base instruction is changed, the model reverts to more rigid behavior. Willison keeps the file under version control to avoid this regression.

What developers should watch next

Three signals will show whether the pattern spreads. First, public releases of new memory features in Claude Code or similar tools will indicate whether other teams adopt the same split. Second, any announced price changes from Anthropic will test whether token routing becomes a standard practice. Third, shared prompt repositories will reveal how many users copy the exact judgment and routing instructions Willison described.

Each of these signals is concrete and measurable within the next three months. If adoption grows, the pressure on fixed prompt templates will increase. If price rises occur first, the incentive to route work to lighter models will strengthen faster.

The overall lesson is narrow but practical. Current models already hold enough judgment to replace some rule sets. Using that judgment while routing routine work to smaller models reduces cost without new infrastructure. The results Willison recorded are one data point in an ongoing shift toward lighter, more selective agent loops.