Most disappointing AI answers were doomed before the model saw them: the goal was vague, the context was missing, and success was undefined. The fix is not a secret phrase.
Talk to it the way you would brief a capable person who just walked in: what you are trying to do, what they need to know, and what done looks like.
Context beats cleverness
The model cannot see your screen, your inbox, or your intent unless you provide them. Paste the email you are replying to. Name the audience. State the constraint: two paragraphs, no jargon, mention the deadline. Ten seconds of context routinely does more than any amount of prompt wizardry.
A useful test: could a sharp person answer well with only what you typed? If they would have to ask what it is for, who is going to read it, and what happened before, then the model needed those things too. The difference is that a person asks, and the model usually just guesses. A vague brief gets you a confident guess at what you meant.
Iterate instead of restarting
The first draft is the opening move. Push back: shorter, warmer, that second point is wrong, use this fact instead. Each correction lands because the whole exchange is in the window. People who get great results are usually three polite demands deeper than people who do not.
Corrections also build on each other. You say shorter, then warmer, then keep the second paragraph, and you end up somewhere you could never have described in one prompt. Nobody can specify taste up front. You converge on it. It helps to treat the model like a fast junior collaborator rather than a vending machine.
Verify at the point of action
You do not need to fact-check every sentence, just the ones that touch reality: the number in the invoice, the date in the contract, the claim in the post you are about to publish. Check where mistakes cost something, and let the rest be a draft.
How much checking an answer needs depends on what a mistake would cost. A brainstorm needs no checking at all. An email to a client needs one careful read before it goes out. A number that is going into a contract needs a real source. Put your attention on the places where an error would actually hurt, and let everything else move fast.