Iterating on output
The first answer is a first draft. Directing revisions, asking the model to critique itself, and knowing when to restart the chat.
Beginners treat the first response as the verdict — accept it or give up. Practitioners treat it as a first draft from a tireless junior: the beginning of a short editing conversation. The model never gets impatient, never resents feedback, and revises in seconds. Use that.
Direct the revision like an editor
- Point at the problem: "The second paragraph buries the deadline — move it to the first sentence."
- Name the delta, not just the wish: "Half the length, keep all three action items" beats "make it punchier."
- Ask for variants when you can't articulate the fix: "Give me three versions: one more formal, one warmer, one blunter." Picking is easier than describing.
- Lock what's good: "Keep the opening exactly as is; only rework the closing ask."
Make the model its own first critic
Before I read this draft: critique it yourself. What are its three weakest points from the perspective of a busy store manager who will skim it in 20 seconds? Then produce a revision fixing those three things.
Self-critique reliably surfaces real weaknesses — models are better editors than they are first-drafters. This one prompt upgrades almost any draft.
When to restart instead
Long editing chats accumulate baggage — old instructions, abandoned directions, your earlier phrasings — and the model keeps trying to honor all of it. If you're five corrections deep and quality is degrading, stop patching. Start a fresh chat and write one new prompt that includes everything you've learned: the facts, the constraints you discovered you cared about, and the best draft so far as your example. Two minutes of consolidation beats twenty of drift.
If three consecutive revisions haven't converged, the problem is upstream: your brief is missing a fact or a constraint. Don't push harder — go back and fix the prompt.