Make AI sound exactly like you (the voice file trick)

If you want to make AI sound like you, stop writing better prompts and start writing one file. The model doesn't know how you talk. The voice file is where you tell it, once.

What you'll have when you're done

• A voice file: one plain-text document with six sections that any AI tool can load
• Proof it works: the same post generated with and without the file, side by side
• A calibration habit, so the file gets sharper every week instead of going stale
• [SCREENSHOT: the finished voice file in a Claude Project]

Before you start

• A Claude account. Free plan is fine. (The same file works in ChatGPT.)
• 5 pieces of your own writing that sound like you: posts, emails, captions, anything.
• 15 minutes.

Step 1: Collect 5 samples that sound like you

Grab 5 pieces of your writing and paste them into one document. Rules for picking:

• Pick pieces people replied to. Replies mean the voice landed.
• Mix formats: a post, an email, a caption. Voice shows up in all of them.
• Skip anything you wrote to sound professional. That's the voice you're trying to escape.

Step 2: Have the AI extract your voice file

Don't write the file by hand. Hand the samples to Claude and make it do the extraction. Paste this in a new chat, then paste your samples where marked:

Here are 5 pieces of my writing that sound like me. Study them, then
build my voice file with exactly these six sections:

1. WHO I AM: one paragraph, first person, based on what the writing
   reveals.
2. WHO I'M TALKING TO: who reads this and what they want from me.
3. VOICE ATTRIBUTES: 3-5 rules that describe how I write, each with a
   one-line explanation.
4. SAY THIS, NOT THAT: 5 pairs. Pull the "say this" side from my actual
   phrasing; write the "not that" side as the generic version I avoid.
5. WORDS I USE / WORDS I NEVER USE: two lists, drawn from the samples.
6. FORMATTING RULES: sentence length, paragraph rhythm, punctuation
   habits, emoji policy. Be specific enough that a stranger could
   imitate me.

Quote my actual phrases as evidence inside each section. Then append
the 5 samples at the end under EXAMPLES, unchanged.

My samples:
[paste your 5 samples here]

This six-section structure is the same one my production voice file uses. The two sections people skip, "say this, not that" and the word lists, are the ones that do the most work.

Step 3: Install it where the writing happens

The file only works if the AI loads it before writing. Two ways:

• Claude Project: create a project called Writing, paste the voice file into the project instructions (or add it as project knowledge). Every chat in that project now starts with your voice loaded.
• Custom GPT: paste it into the GPT's instructions.

Step 4: Run the A/B test

Prove the file works before you trust it. In your writing project, send:

Write a 150-word post announcing [your topic] two ways.
Version A: ignore my voice file completely. Write it how you'd
write for anyone.
Version B: follow my voice file exactly.
Label both. Don't blend them.

Version A is what everyone else is posting. Version B should read like your best writing day. If you can't tell them apart, your samples in step 1 were too generic; swap two of them for spikier pieces and re-run step 2.

Step 5: Calibrate when it drifts

The file is never done, it's maintained. When a draft feels off, don't fix the draft. Fix the file:

Version B is close but not me. Notes: [what felt off, e.g. "too salesy,
I never use exclamation points, my sentences are shorter than this"].
Update the voice file so this never happens again, and show me the
full updated file.

Paste the updated file back into your project instructions. Two or three rounds of this and the drift mostly stops.