source: simon willison: fable's judgement
level: technical
simon willison learned from a fireside chat with the claude code team that it is better to let fable use its own judgement rather than giving it strict rules. for example, instead of telling fable to only write tests for large features, you should tell it to decide for itself when tests are needed. this approach leads to better results because the model can assess context more flexibly than a rigid instruction.
another tip from jesse vincent helps conserve fable tokens, especially with upcoming price increases. you can tell fable to delegate coding tasks to subagents running cheaper models, using its own judgement to pick the right model for each task. willison prompted claude code to use a lower-power model for coding tasks, and it saved a memory file to guide future behavior. the memory specifies that implementation work can use sonnet or haiku, while judgement-heavy tasks stay with the main model.
this delegation strategy is working well for willison. he reports getting a lot done while using fewer fable tokens. the approach separates routine coding from high-level design and review, keeping the main model focused on synthesis and decision-making. by trusting the agent's judgement, developers can reduce costs without sacrificing quality.
why it matters: trusting ai agents to make context-aware decisions about testing and model selection can reduce costs and improve efficiency in software development workflows.