How to Keep AI-Drafted Emails and Messages From Sounding Like AI

AI drafts of emails, Slack messages and LinkedIn replies have a look. People spot it fast, and then read everything else you wrote with one eyebrow up. Here's how I keep drafted correspondence sounding like a person: the tells to cut, a style block to hand the model, and a checklist to run before sending.

The goal is narrow. It isn't about gaming a detection tool. It's whether the message reads like someone wrote it in one sitting.

Why short messages give it away

An essay can bury a few machine habits. A five-sentence email can't. These are the patterns that make a reader suspicious:

• Scaffolding that says nothing. "I hope this email finds you well." "I wanted to reach out regarding." "Please don't hesitate to reach out." Start with the first real thing you'd say. End when you're done.
• Structure the length doesn't need. Bullets, bold labels and numbered steps inside what's really three sentences. A short message is prose.
• Over-politeness. "Thank you so much for taking the time to..." One plain thanks is warmer than three inflated ones. "Happy to" beats "I'd be more than happy to."
• Restating before answering. Just answer.
• Empty validation. "Great question." "Absolutely!" Skip it and respond to the substance.
• The summary sentence. A closing line that restates the message with a bow on top.
• Triples on autopilot. "Clear, concise, and compelling." Real speech doesn't come in tidy threes.
• Stacked hedging. "I think it might potentially be worth considering." Pick a position.
• The em dash tic. Models reach for it constantly. One in a short note is fine. Three is a signature.
• The wrong register. Stiff with a close colleague, chummy with a stranger.

LinkedIn adds its own cadence: one-line paragraphs stacked for drama, "It's not just X. It's Y.", a manufactured arc from small vulnerability to tidy lesson, and a "Thoughts?" glued to the end. The fix is the same. Say the actual thing, with one concrete detail. The polished arc reads as content marketing.

The prompt

Paste this at the top of the chat, or save it as a custom instruction or project instruction, before you ask for the draft. It's the condensed version of the style rules I use.

You're drafting short correspondence (an email, a chat message, a LinkedIn reply) that a real person will send under their own name. It should read like they wrote it in one sitting.

Do: use clear, simple language. Start with the actual first thing to say. Use contractions. Mix short sentences with longer ones. Use active voice. Take a position instead of hedging. Use specific, plain nouns and concrete details. Address the reader as "you" where it fits. Match the formality to the relationship.

Don't: open with "I hope this finds you well" or "I wanted to reach out." Close with "Don't hesitate to reach out" or "I look forward to hearing from you." Restate the question before answering it. Open with praise like "Great question." Add a closing sentence that summarizes what you already said. Use bullets or bold in a message that's only a few sentences long. Write in tidy lists of three by habit. Stack hedges. Use em dashes. Use marketing language, metaphors or cliches. Use words like utilize, tapestry, realm, pivotal, harness or testament.

No preamble or notes. Give me only the message.

How to use it

Give the model the facts first: who it's to, how well you know them, and what you need to say. The style block handles how it sounds. It can't supply what you mean.

Then treat what comes back as a draft. The checklist below catches what slips through.

Tip

If you've got a few short messages you've written yourself, paste two or three in as samples and say "match this voice." A handful of real examples steers tone better than any list of rules.

The checklist

A 30-second pass before you send.

Cut on sight

• "I hope this email finds you well" / "I hope you're doing well"
• "I wanted to reach out to..." (just say the thing)
• "Please don't hesitate to reach out" / "Feel free to reach out with any questions"
• "I look forward to hearing from you," unless you mean it as a nudge
• "Thank you so much for taking the time to..." (make it "Thanks for...")
• "I'd be more than happy to..." (make it "Happy to...")
• "That's a great question" / "Great point" / "Absolutely!"
• Any sentence that restates their question before answering it
• Any closing sentence that re-summarizes what you already said

Check the shape

• Bold, bullets or numbered steps inside a three or four sentence note? Turn it back into prose.
• Are the sentences all about the same length? Break one up or run two together.
• Count the em dashes. More than one in a short message is a tell. Swap most for commas or periods.
• A "clear, concise, and compelling" style triple? Cut the third item, or the whole pattern.

Check the voice

• Contractions everywhere they'd naturally go?
• Hedged three times when once, or zero, would do? Pick a position.
• Does the formality match the relationship?
• Read the first line out loud. Would you really start a message that way?

Before it leaves your machine

• No local file paths, folder names or filenames in the body. Those are for you, not the reader.

What it's good at, and where it isn't

This fixes the surface. It won't make a message say something it doesn't. If the draft is vague, cutting the filler just makes it shorter and vague.

It also isn't a disguise. Plenty of people draft with AI, and that's fine. A message with your name on it should sound like you and say what you mean. Do both and nobody cares how the first draft got written.

Here's the difference in miniature.

Hi Sam, I hope this email finds you well. I wanted to reach out regarding the delivery delay. I truly appreciate your patience. Please don't hesitate to reach out with any questions.

Hi Sam, the delivery slipped to Thursday because the parts shipped late. I'll confirm once it's on the truck. Anything you need before then?

The second one is shorter, says more, and sounds like someone who knows what's going on. That's the whole game.

For the other direction, making the model's own answers shorter and flatter, see Execution Mode.