Every coach and consultant I talk to has typed some version of this: "Write this in my voice. I'm warm, direct, a little funny, and I don't like fluff."
And every one of them got back something warm, direct, a little funny, with no fluff, that sounded like nobody in particular.
That's not a model problem. It's that the instruction is empty. If you want to know how to make AI sound like you, the first step is to stop describing your voice and start showing it. Here is what that looks like, and the short version you can do this afternoon.
TLDR
- "Write in my voice" plus 4 adjectives describes about 40 million people. The model picks the average of them.
- A Voice Print is not a description. It's evidence: your rhythm, your banned words, your actual sentences.
- 3 questions replaced my entire voice description, and the output got better the same week.
- The 20-minute version needs 1 recording, 1 email, and 1 thing you sell.
- This stops working at exactly 1 point, and it's not where you'd expect.
Why "Write in My Voice" Is an Empty Instruction
Think about what the model actually receives. "Warm." "Direct." "A little funny." Those are not properties of your writing. They're properties of most writing that people like. The model has seen millions of pieces described exactly that way, and it gives you their average.
I've made the argument before that generic output is a context problem, not a model problem. This is the most common shape of that problem. You gave it a personality profile. It needed a transcript.
Here's a test. Write down 4 adjectives for your voice. Now write down 4 for a coach you admire. If 3 of them overlap, and they will, the adjectives were never going to get the model to you.
What a Voice Print Actually Contains
A Voice Print is what I build in the Encode step of the method, and the thing to notice is that it contains almost no adjectives. It contains evidence:
- Rhythm. Your average sentence length, and how far you swing from it. Some people write long then land a 2-word sentence. Some never do. That's more distinctive than "direct."
- The explanations you've refined. Every coach has 3 to 5 things they've explained 200 times out loud. The 200th version is polished in a way no fresh draft can be. Those go in verbatim.
- Banned words. Not the generic AI-tell list. Yours. The compliment you find embarrassing. The word your industry overuses that you refuse to touch. A system that knows what you reject holds a line that a system that only knows what you like will drift across.
- Structural habits. Do you open with the point or with a scene? Do you use lists or hate them? Do you end with a question or with a flat statement? Those are rules, and they're yours.
- Register shifts. How you sound in a client email versus a post versus a sales call. A voice is not 1 setting. It's a range, and the model needs the edges of the range, not the middle.
None of that comes from you describing yourself. All of it comes from your material. That's why the 3 Sources exist: what you say, what you write, what you sell. 10 to 20 pieces across those 3 is enough to build a Voice Print that holds.
The 3 Questions That Replaced My Voice Description
For a long time my own skills carried a paragraph describing the voice they were supposed to produce. Then I replaced it with 3 questions the draft gets read back through, and honestly the output improved the week I made the swap.
For restaurant copy, the 3 questions are: Does this line mean anything, or is it poetry that means nothing? Is this a resume or a history lesson? Does this make me want to go and eat there?
For a coach, the 3 might be: Would I say this sentence out loud on a call? Is this something I've actually seen with a client, or something I read? Would my best client recognize me in this paragraph?
The point is that questions are checkable and adjectives are not. "Warm" cannot fail a test. "Would I say this out loud" can, and does, and every failure teaches the system something.
Do This Instead: The 20-Minute Version
You don't need the full build to see the difference. Here's the version I'd do if I had 20 minutes and a chat window.
- Minutes 1 to 5. Find 1 recording where you were explaining something to a client. Transcribe 5 minutes of it. Don't clean it up.
- Minutes 6 to 10. Find 1 email or message you wrote to a client when you were annoyed or excited. Unpolished. The real reply, not the newsletter.
- Minutes 11 to 13. Grab 1 paragraph from something you sell. A module intro, a sales page section, the onboarding email.
- Minutes 14 to 17. Write 5 words or phrases you would never use, and 3 sentences of yours that you actually like. 1 line each on why.
- Minutes 18 to 20. Give the model all of it and say: "Here is how I actually talk and write. Match the rhythm and the vocabulary in these, not a description of them. Before you show me the draft, check it against the 3 sentences I like and the 5 phrases I never use."
That's a tiny Voice Print with a tiny verification loop. It will beat "write in my voice, I'm warm and direct" every single time. It is not a system, and I'd be lying if I said it holds up across 50 pieces. But it will show you, in 1 draft, that the problem was never the model.
Where This Stops Working
Two places, and neither is where people expect.
It stops working when the material goes stale. A Voice Print built in March describes March you. Your offer changed, your framework got a new name, you stopped saying that one thing. Nothing in a chat window knows that. A real system has a freshness check. The 20-minute version doesn't, so redo it every couple of months.
And it stops working on taste. A coach I work with told me, about my line-by-line edits on a draft: "Even if I agreed with 100% of your feedback, I don't think you could edit the copy the same way I would." He's right. A Voice Print gets you to a draft that sounds like you. The last 10%, the part where you'd have chosen this word and not that one, is still yours. I don't have a rule file for taste, and I'm not sure anyone does. I've written about that in the thing AI can't replicate.
So do the 20 minutes. Then decide whether the 90% it gets you is worth building properly.
If you've read this far and you're still going to type "write in my voice" one more time, I can't help you.
Kidding. That's exactly who I work with. Send me 1 piece of your work, something you said, wrote, or sell, and you get a finished asset back in your voice within 48 hours. It'll be the first time an AI draft sounds like you instead of like a description of you.
FAQs about making AI sound like you
How do I make AI sound like me?
Stop describing your voice and start supplying it. Give the model real samples of how you talk, how you write to clients, and what you sell, plus a short list of words you never use and sentences you like. Then have it check the draft against those samples before showing you. Descriptions like "warm and direct" produce the average of everyone described that way.
What is a Voice Print?
A written record of your actual writing and speaking patterns, built from your own material rather than from adjectives. It captures sentence rhythm, your refined explanations, banned words, structural habits, and how your register shifts between a post, an email, and a call. It is the first output of the Encode step and the foundation of a reference library.
How much material do I need?
For a quick test, 1 recording excerpt, 1 unpolished email, and 1 paragraph from your paid offer. For a system that holds up across months of output, 10 to 20 pieces across the 3 Sources. Volume past that adds very little. Range across sources adds a lot.
Why does AI still sound slightly off even with samples?
Usually 1 of 2 reasons. The samples are stale and describe an older version of you, which a freshness check fixes. Or you are hitting the taste layer, the final 10% of word choices that no reference library fully captures. A good system gets you to a draft you recognize. The last pass is still yours.