How to make AI write in your voice, not "like AI"

• Threads
• Method
• ≈ 7 minutes

Why AI texts get spotted from the first line — and what to do about it.

AI-generated text isn't recognized by meaning, but by tells: even rhythm, symmetric constructions, and words you'd never use in real speech. Tells can be described, which means they can be banned.

People say "it sounds like AI" but usually can't explain why. We broke it down on our own texts and got a list of tells.

One tell stands apart — the em dash. A machine uses it far more often than normal, because it glues any two thoughts together without grammar.

The setup splits into two, and you shouldn't mix them up. The first layer is what's banned outright: the tells above are the same for any author.

The second layer is who you are: sentence length, how you address the reader, whether jokes appear or not, favorite turns of phrase. This one is unique to each person.

One layer without the other doesn't work: bans alone give you sterile text, voice without bans gives you recognizably machine-made text.

The profile is captured from what you've already written. Posts, client emails, messages in a work chat all work — anything where you wrote it yourself and didn't try too hard.

Five texts give a blurry profile, ten give a precise one. More than ten adds almost nothing.

Act as an editor. Here are 5–10 of my texts: [paste texts]

Output a profile of my voice as a list:
1. Average sentence and paragraph length, counted in words.
2. How I address the reader: formal or informal, capitalized or not.
3. Whether there are jokes and who they're aimed at: myself, the subject, the market.
4. 10–15 phrases I use more often than others.
5. What's NOT in my texts: words and constructions that aren't like me.
6. Punctuation: exclamation marks, ellipses, dashes — how many and where.

Return this as a compact list I'll plug into other prompts.

There are two more things from Anthropic's documentation that apply directly to a voice profile.

For a voice profile, this means the following. Your ten texts go right at the start, the profile and bans follow in a separate tag, and the actual task comes last.

Save the profile you get. From there you have two paths, and you need to choose deliberately.

Claude's memory across conversations works and is on by default, but each project has its own separate memory. So it's more reliable to set up a dedicated project for your blog and keep the profile there: your other conversations won't leak into it.

If you're writing in an incognito chat or a different model, plug the profile in manually with every prompt. Incognito mode isn't saved to memory at all — that's the whole point of it.

How this profile applies to someone else's format — in the breakdown on repacking posts that took off.

Reading aloud catches the rhythm. Wherever you ran out of breath mid-sentence, there should be a period; wherever it's smooth as a metronome, it's missing some roughness.

The second step is tougher. Strip the topic out of the text and see what's left: phrase length, tone of address, intonation. If what remains could belong to any author in your niche, the voice didn't work.

You borrow the mechanics, you put in your own meaning. It took us years to get this. You get it in seven minutes of reading.

First day free, no card, no sign-up.

Not ready for the tool yet — the prompt above works and it's free.