Break work into small iterative chunks

Never ask an agent for a monolithic output. One function or one task at a time prevents the model from producing a jumbled mess. Large leaps create inconsistency and duplication. Each chunk should be small enough to handle within context and for you to understand.

Large monolithic asks produce inconsistent, duplicated code because the agent loses track of earlier decisions within the same response. By the time it reaches step five, it has forgotten the pattern it established in step two. Small chunks prevent this: each task is self-contained, tested before (see: spec before code for better agent output) moving on, and committed as a checkpoint. This also matches how senior engineers decompose work naturally; the agent benefits from the same structure that makes human code review manageable.