Your AI Writing Still Sounds Like AI (here's why)...
The prompt that capture's your real voice, not the one you are trying to be.
For a long time I thought capturing your voice for AI writing meant feeding it your old work.
Paste in twenty newsletters. Ask it to find the patterns. Build a style guide.
I tried it with multiple tools. The output was always the same kind of useless. Generic descriptions of my voice that read like a LinkedIn bio.
“Direct, no-nonsense, accessible.”
Sure. So is everyone.
The problem is your published writing is already the wrong input. By the time you publish something, you’ve edited the life out of it. You’ve cleaned up the asides. Smoothed out the rhythm. Picked the safer word over the truer one.
The AI reads the polished version and learns to write polished. What you actually wanted was the version BEFORE the polish.
I figured this out by accident.
I have been using Wispr Flow to dictate everything. Substack notes. Slack messages. Half my client emails. One day I noticed the dictated stuff sounded more like me than the typed stuff.
I wasn’t being more careful, I actually wasn’t being careful at all. I was just talking. The voice was already there. I just had to stop translating it into typing.
That’s when I tried something different.
Instead of feeding Claude my published work, I had Claude interview me. Six prompts across different modes (opinion, story, teaching, a few others). I voice-to-texted my answers. Raw. No editing. Claude analyzed each answer, named the patterns it was seeing, and built a voice profile from the actual way I talk.
It worked better than anything I’d tried before. Not by a small margin.
The voice profile that came out of that conversation could shape my messy thoughts into things that sounded like me to me. That’s the bar. Not “sounds like a confident newsletter writer.” Sounds like the specific person who wrote it.
How the Interview Works
The structure is a conversation, not a form. The AI runs you through five or six questions, each pulling a different mode of your voice out of you.
You talk differently when you’re explaining something than when you’re annoyed. You talk differently when you’re telling a story than when you’re admitting you were wrong. The questions cover all of those registers so the final voice profile knows which mode to apply for any given piece.
The modes the AI tries to extract include:
Opinion mode. How you argue and how you build a case for a position.
Change-of-mind mode. How you handle uncertainty and how you frame your own evolution on a topic.
Teaching mode. How you instruct someone through something you know cold.
Annoyance mode. How you criticize patterns and what your structural diagnosis looks like.
Story mode. How you build narrative and where your humor actually lives (which is almost never where you think it is).
Short-form take. How you compress a thought without losing voice.
Between each answer, the AI does the analysis. Names what it heard. Flags the moves you made that you weren’t aware of. You correct it where it’s wrong.
By the end of the interview, the AI has enough material to build a voice document that reflects how you actually talk in each mode.
Why Voice-to-Text Is the Unlock
Speed is the first reason.
You can talk through a 300-word answer in two minutes. Typing the same answer takes ten and your brain edits the whole way. Voice-to-text removes the editing layer between thought and text.
The second reason is no censoring.
When you type, you watch your words appear and you fix them. When you talk, the words are gone before you can second-guess them. The asides slip out. The honest small admissions slip out. The way you actually phrase something instead of the way you’d phrase it in writing slips out.
That’s the data the AI needs.
The third reason is rhythm.
Your spoken sentences have rhythm. They breathe. They land. Your typed sentences are often shorter and flatter because you’re constructing them word by word instead of releasing them as full thoughts. The AI learns better rhythm from your speech than from your typing.
I use Wispr Flow specifically because the cleanup is good enough that you don’t have to fight it. It catches paragraph breaks. Handles punctuation. The model behind it actually knows when you’re rambling versus when you’re listing.
If you want to try it, my affiliate link gets you a free month of pro. I’d recommend it whether you used my link or not. It’s the tool that made this whole approach possible for me.
What You Do With the Voice Profile
Once you have the profile, you use it as a system prompt. But not the way most people use AI for writing.
You don’t ask it to write a piece from scratch. You write the messy version yourself first.
Voice-to-text a rough dump of what you want to say. Or type out a brain-dump in whatever order it comes. Don’t worry about structure. Don’t worry about polish. Get the substance out of your head and into a doc.
Then you paste your messy version into Claude (or whichever AI you use) with the voice profile loaded. The AI takes your raw material and shapes it into your voice, in the right mode for the piece, with your structural patterns applied and your anti-patterns flagged.
What comes back is already yours. The thinking is yours. The substance is yours. The voice is yours. The AI did the polishing.
This is where most people stop short. They build a voice profile and then keep using AI to generate writing from a topic prompt. That’s still going to give you AI writing. The voice profile only works if you give the AI your actual material to work with.
I’ve been running this on every piece I write for the last two weeks. Newsletter articles. Substack notes. Client work. The first draft I get back is in my voice because I started from my own thinking, not from “write me a newsletter about X.”
I still edit. The polish isn’t perfect. But I’m editing my own polished voice instead of fighting AI sludge to sound like me. That’s a different kind of work and a much faster one.
The Voice Extraction Prompt
Paid subscribers get every prompt and workflow I publish like this one. The full voice extraction system. The Cowork setup I use to run the whole thing. The skills I layer on top to handle different platforms (newsletter, X, notes). If you’re tired of fighting your AI drafts to sound like you, this is the system that fixes it. Upgrade link below.
