The real reason your AI drafts sound wrong isn't the prompt
Most people trying to fix bad AI output do it by rewriting the prompt five different ways. That's backwards.
Every draft AI produces is downstream of four decisions you made (or didn't make) before you ever typed anything:
Who is this actually for — not "customers," but the specific person reading it.
What should happen after they read it — a reply, a click, a signed form, nothing.
What format does it need to take — email vs. internal doc vs. social post are not interchangeable.
What constraints does it have to respect — length, tone, facts that can't be fudged, things you're legally or reputationally not allowed to say.
If you skip these and go straight to "write me an email about X," you're asking the model to guess all four — and then you're stuck rewriting the prompt to patch a guess, over and over.
The fix isn't a better prompt. It's answering those four things on paper first, in under a minute, before you touch AI at all. Once the task is actually defined, almost any reasonable prompt produces a usable draft — because the thinking already happened.
The other habit worth building: don't approve/reject AI drafts as one binary decision. Sort them into three buckets before they go anywhere near a customer —
🟢 safe as-is
🟡 needs a fact-check
🔴 needs a tone edit
Most people either rubber-stamp everything or rewrite everything from scratch. Neither scales. Sorting first tells you exactly where your 2 minutes of editing time should go.
None of this requires prompt engineering. It requires defining the task before you write the draft.