How to Prompt AI for Writing (That Doesn't Sound Like AI)

AI is a strong writer and a weak stylist. Out of the box it produces the statistical average of everything ever written on your topic — competent, forgettable, and full of tells like “in today's fast-paced world” and “unlock.” The fix isn't a better model. It's a better prompt.

Why AI writing sounds like AI

A language model predicts the most likely next word. “Most likely” is exactly what makes writing generic — the safest, blandest phrasing wins by default. When your prompt is vague (“write a blog post about productivity”), you're asking for the average of a million productivity posts. You'll get it.

Every technique below is really one idea: give the model enough specific constraint that the “average” answer is no longer an option.

Give it a voice, not just a task

The single highest-leverage thing you can do is show the model how you write. Adjectives (“write it casually”) barely move the output. Real examples move it a lot.

Also tell it who's reading. “Explain our pricing” and “explain our pricing to a skeptical CFO who thinks we're overpriced” produce completely different drafts. The reader shapes the voice more than any style adjective.

Constrain hard: length, format, and what to avoid

Models pad. Left unconstrained they'll give you a five-paragraph answer to a one-line question, three throat-clearing sentences before the point, and a “In conclusion” you didn't ask for. Constrain all of it up front:

• Length: “Under 120 words.” “Exactly three sentences.”
• Structure: “Hook, then three short sections with subheads, then one takeaway.”
• A ban list: the most underrated instruction in AI writing.

Show, don't tell

“Make it punchy” is a wish. A before-and-after example is an instruction. When you want a specific quality, demonstrate it:

One good example teaches the model more than a paragraph of description. This is also the fastest way to fix a draft that's almost right: paste the sentence you like and the sentence you don't, and say “more like the first.”

Write in passes, not one shot

Good writers don't draft and publish in one motion, and neither should your prompt. Separate the jobs:

• Pass 1 — structure: “Give me three different angles and a one-line outline for each. Don't write the piece yet.”
• Pass 2 — draft: pick an angle, then “write the draft in my voice.”
• Pass 3 — cut: “Cut this by 30% without losing meaning. Kill every sentence that doesn't earn its place.”
• Pass 4 — sharpen: “Make the first line impossible to scroll past.”

Asking for options before committing (Pass 1) is the step most people skip, and it's the one that avoids the generic default answer.

Let a second model edit

Every model has a house style. Ask the same model that wrote a draft to critique it and it will mostly agree with itself. A different model catches what the first one is blind to — the cliché it loves, the rhythm it defaults to, the claim it waved through.

A reliable two-step: have one model draft in your voice, then hand the draft to another with “You're a sharp editor. List what's weak — clichés, filler, anything that sounds AI-generated — then give me a tightened version.” This is exactly the kind of draft-then-critique loop AskOnce runs across models for you, so you're not copy-pasting between tabs.