Task → do → verify

The single biggest difference between preparers who get magic from AI and preparers who get slop: the loop. Give it a task, let it do the task, give it a way to verify its own result. For a notice-explanation drafter, verification means checking the draft against the de-identified notice summary AND against a rule list (no figure the notice didn't carry, every authority VERIFY, nothing identifiable) before you ever see it. And one good run proves less than it seems. The model is non-deterministic: the same prompt on the same files can come back different next time. So the check has to run every time, not once when you first set it up.