Breaking Big Tasks Into Agent-Sized Steps

Vague, large prompts produce vague, large diffs — and vague diffs are the hardest kind to review. The fix isn’t a cleverer prompt; it’s decomposition.

Signs a task is too big

• You can’t describe what “done” looks like in one sentence
• The task touches more than two or three unrelated parts of the system
• You’d need to check in multiple times even if a human were doing it

If any of these are true, break it down before prompting.

A decomposition pattern that works

1. State the end goal in one sentence, even if the path there is multi-step.
2. List the components the goal actually requires — data layer, API, UI, tests — and order them by dependency.
3. Hand the agent one component at a time, reviewing before moving to the next.
4. Reassess after each step — the plan for step 3 often changes based on what step 1 revealed.

This is slower than firing off one giant prompt and hoping. It is also the difference between a working feature and an afternoon spent untangling a diff that touched twelve files for the wrong reasons.

That wraps the “Prompting Patterns” series for now — more installments coming as new patterns prove out in practice.