1. The first answer is an opening bid

Skilled users rarely accept output one. Not because models are bad, but because the first answer reveals what the model assumed, and now you can correct the assumptions. 'Good, but the reader is more senior than this' or 'right structure, wrong emphasis, lead with cost' are second prompts that do real work.

Iterating on a draft preserves what is already right. Re-rolling from scratch throws away the 70% that was fine along with the 30% that wasn't.

2. Make the model critique itself

Before you even read closely: 'List the three weakest parts of this draft and why.' Models are genuinely useful critics of their own output because critiquing is a different task than writing, different framing, different result.

Then choose: fix all three, fix one, or overrule the critique. You stay the editor; the model does editor's-assistant work.

3. Let it ask you questions

For any complex task, add one line: 'Before starting, ask me the questions whose answers would most change the result.' Three sharp questions later, you have surfaced the context you forgot you were assuming.

This inverts the usual failure: instead of discovering the missing context after a wrong draft, you pay ten seconds up front. It is the cheapest quality lever in this entire course.

4. Iterate vs restart

Iterate when the skeleton is right and the flesh is wrong: tone, emphasis, length, examples. Restart when the frame itself was wrong: wrong task, wrong audience, or a conversation so long the history is fighting you.

A restart is not starting over: you carry forward the anchor summary, the constraints that proved necessary, and the critique of the failed attempt. That is the difference between rolling dice and running a loop.