Lesson 2
How to Teach AI to Write Like You, With One Voice File
Give AI an interview prompt and answer its ten questions out loud. It writes a voice file from your own words. Then one line in CLAUDE.md, the file AI reads first, sends it there before it writes.
Set it up once, in one sitting
Published 4 November 2026
By Manie Eksteen
In this lesson
- Teach AI to write like you: what you need
- Step 1: Open a new chat in your project
- Step 2: Paste the interview prompt
- Step 3: Answer out loud
- Step 4: Check the file, then save it
- Step 5: Point CLAUDE.md at your voice file
- How your voice file grows into a folder
- Voice files for a team
- Using ChatGPT instead?
- Frequently asked questions
- The takeaway

If you're still rewriting every draft AI gives you, it's because it doesn't know what you sound like yet. Here's how to teach AI to write like you and sound like you, and it all starts with one file: voice.md.
A voice file is a short document about how you write: how you greet people, how you say no, the words you always use and the ones you never would, with a few examples in your own words. Give AI that file before it drafts anything, and the draft starts from you. Which means you rewrite far less.
Our own office is set up this way: every desk reads our voice file before it writes. This lesson gets you the same thing in one sitting. AI interviews you, you answer out loud, and it writes the file from your answers.
The steps are checked against Anthropic's, OpenAI's and Apple's own pages on 3 November 2026; where CLAUDE.md sits comes from lesson 1. The sources are listed at the end of this lesson.
Teach AI to write like you: what you need
- The Claude desktop app, with a project linked to a folder on your computer. If you haven't set that up, lesson 1 shows you how, and it takes one sitting. It also made you a starter voice file, which this lesson fills in.
- A microphone: the one in your laptop is fine.
- Ten questions' worth of time, somewhere you can talk.
Using ChatGPT? The same prompt works there; the changes are at the end of this lesson.
Step 1: Open a new chat in your project
- Open the Claude desktop app. Go to the project you made in lesson 1.
- Start a new chat inside that project. A chat in the project can work with the project's folder, so the file it writes lands there.
Keep the app open until the file is saved. Depending on your settings, Claude may ask before it saves the file: say yes.
Step 2: Paste the interview prompt
Copy this into the chat and send it. It works for any business, so you don't need to change a word.
Interview me so you can write like me. Ask me ten questions, one at a time, and wait for my answer before the next. I'll answer out loud, so keep my words as they come. Ask about: how I greet a customer, how I answer a question about what we offer, how I say no, how I say sorry, how I write a post, the words I always use, the words I never use, how formal I am, a story I tell about the business, and how I sign off. When we're done, write my voice file as voice/voice.md: how I sound, the words I use and never use, and three short examples in my own words. Then show it to me before you save it.
Set it up and hit a bump in the road? Email me at manie@tidastra.ai and I'll get back to you with how to fix it, free of charge.
Step 3: Answer out loud
- Click the microphone in the chat box. The arrow beside it switches between dictation, which types what you say, and voice mode, a spoken conversation.
- Answer each question the way you'd say it to a guest or a customer. Don't tidy it up. The way you really talk is the point.
- Send each answer and wait for the next question. It asks all ten, one at a time.
Why out loud? When you type, you tidy yourself up. When you talk, everything in between comes through, so AI gets a clear idea of how you really sound. Claude reads the words your dictation types, not the sound of your voice.
On a Mac you can also switch on the Mac's own dictation and use it in any text box, if you'd rather not use the app's microphone. Or use a dictation app. Wispr Flow works too, and here's the free one I use on my Mac.
Step 4: Check the file, then save it
- Read the voice file it shows you. It should sound like you on a good day.
- Correct anything that doesn't. Tell it in plain words: "I'd never say that", "I'm more direct than this".
- Say yes when it's right. It saves the file as
voice/voice.mdin your project's folder. If you did lesson 1, that's the starter file, now filled in.
Step 5: Point CLAUDE.md at your voice file
The file sitting in your folder isn't enough on its own. Claude can open it, but there's no promise it reads it before it writes. Your CLAUDE.md is what sends it there: it's the first file Claude reads in every chat (lesson 1's instructions make sure of that), and its map already lists voice/voice.md.
- Open CLAUDE.md at the top of your folder.
- Find the voice file's line in the map. From lesson 1, it says to open the file before writing a reply to a guest.
- Change it so it covers everything you write:
The line can live in two other places. If you use skills, make the first step of any skill that writes, say an email, "read voice/voice.md" in your business folder. And if you've made a project just for writing, say one for emails, that project's instructions can carry it.
Then test it. Ask for something you write every week, like a reply to an enquiry or a short post, and read it against how you'd write it yourself. If a line still sounds off, tell Claude what's wrong and ask it to add the fix to your voice file, so the next draft comes out closer.
How your voice file grows into a folder
One file is all you need to start. In time, though, you'll notice the way you write and the way you speak are different. A reply to a guest isn't how you'd say it on camera or in a voice note.
So the one file becomes a folder:
voice.md, how you writespoken.md, how you speak, for scripts and voice notes- a folder of examples: your best replies, posts and emails
INDEX.mdat the top, which says which file is for what
When you get there, change the voice line in CLAUDE.md's map to this:
Then change voice/voice.md to voice/INDEX.md in any skill or writing project that points at your voice file.
Voice files for a team
If you run a team, each person gets their own voice file, in their own folder inside voice/. Each person does the same interview, answering out loud as a raw voice note, so their file sounds like them and not like the business's polished version.
A project linked to a folder on your computer lives on that computer, so it isn't shared. Share the folder's files instead, and each person points their own project at them. When AI writes on someone's behalf, it reads that person's file.
The version we build for companies goes further: a voice folder that can hold up to a hundred files of examples, and a checker that reads them all against the finished draft and sends back a few fixes. That way the AI doing the writing reads one or two files, not every example.
Using ChatGPT instead?
The interview prompt works the same. Three changes:
- Answer with voice in ChatGPT, or with your computer's dictation.
- In the ChatGPT desktop app, a local project can attach a folder on your computer, and ChatGPT can read and change the files in it, so it can save the voice file there. On the web, save the finished file to your project instead: open the response's menu and choose Add to project sources.
- In the desktop app, the map file is AGENTS.md, the name OpenAI's tools look for, as lesson 1 explains: add the voice file's line there. In a project on the web, there's no folder: add the voice file to the project's sources, then put this in the project's instructions (the menu, then Project settings): "Before you write anything for me, read my voice file in this project's sources and write in that voice."
Frequently asked questions
What is a voice file?
A short document that tells AI how you write: your tone, the words you use and never use, how you greet people and sign off, with a few examples in your own words. It's a brand voice guide, written for AI to read before it drafts anything.
Do I have to answer out loud?
No, typing works too. Out loud is better because you don't edit yourself while you talk, so your real phrases and rhythm end up in the file. Claude reads the words your dictation types, not the recording.
Will AI read my voice file on its own?
Don't count on it. Claude can open any file in a project's folder, but there's no promise it reads one before it writes. The voice file's line in CLAUDE.md, the first file Claude reads, is what sends it there first.
Can I use one voice file for my whole team?
You can start with one for the business, but each person writes differently. For a team, give each person their own voice file from the same interview, and have AI read the right one for whoever it's writing for. If you'd like it set up for your whole team, book a call.
How do I keep it up to date?
Whenever a draft sounds wrong, tell Claude what was off and ask it to add the fix to your voice file. Each correction makes the next draft closer, and the file grows with you.
The takeaway
AI can only write like you once it knows how you sound. Ten questions, answered out loud, give it your voice in one file, and one line in CLAUDE.md sends AI to that file before it writes. Start with the one file today. It grows into a folder when you need it to.
Sources, read on 3 November 2026
Finished this lesson?
Mark it done and it goes into your kit: a voice file AI writes from.
