🪞 Teach AI To Write Like You (Not Like A Press Release)

AI’s default writing voice is a press release. There’s a reason for that: it learned to write from the internet, and the internet’s most polished, most public writing is corporate - announcements, landing pages, LinkedIn posts by people “thrilled to share”. So when you ask it to “write an email”, it reaches for its poshest register, and out comes something you would never, ever say.

But here’s the thing nobody tells beginners: your actual voice already exists, in writing, in enormous quantities. It’s sitting in your sent folder and your WhatsApp threads. You’ve written thousands of messages in your natural voice without once thinking about “tone”. So instead of describing your voice to AI (nearly impossible - try describing your own accent), you show it the evidence and let it work your voice out.

I call this the Three-Message Mirror, and it takes about fifteen minutes. By the end, Claude will hold a written description of how you write - specific enough that the messages it drafts sound like you on a good day, not like a brand having a soft launch.

Step 1: Choose your three messages

What it is: three real things you’ve written, at normal speed, to real people. Not your best writing. Your normal writing.

How to do it: go digging in your sent folder and your messages, and pull out:

1. One quick message - a WhatsApp or text, a few lines, to someone you’re relaxed with
2. One everyday email - to a client, a colleague, the school office, whoever you email routinely
3. One longer thing - a proper email, a post, a message where you explained something or handled something delicate

The bit everyone gets wrong: picking their most impressive writing. Don’t. The mirror should reflect Tuesday-afternoon-you, not job-application-you. If you feed it your fanciest work, you’ll get an AI that impersonates you showing off, which is its own kind of wrong. Boring messages are perfect. Boring messages are the good ones.

One privacy note before you paste: swap out anything sensitive - other people’s surnames, account details, anything you wouldn’t want on a sticky note. Your voice survives the redaction fine.

Step 2: Run the mirror prompt

What it is: the analysis. You hand over the three messages and ask Claude to study how you write, not what you wrote about.

How to do it: open a fresh chat and paste this, with your three messages filled in:

Nick this one, fill in the blanks:

I'm going to paste three real messages I've written to different people. Study how I actually write - ignore the topics, focus on the voice. Look at: how I open and close, my sentence lengths, how formal or casual I am, words and phrases I reuse, how I soften things (or don't), my punctuation habits, and how I sign off. Notice how my voice shifts depending on who I'm talking to.

Then write me a short "How [YOUR NAME] writes" note I can save - specific enough that you could draft a new message in my voice and someone who knows me wouldn't blink. Quote short examples from my messages as evidence for each observation, and include a section on how I adjust for different audiences.

Message one (a quick one, to [who]): [PASTE]
Message two (an everyday email, to [who]): [PASTE]
Message three (a longer one, to [who]): [PASTE]

The bit everyone gets wrong: skipping the “to [who]” labels. Your voice isn’t one voice - you write differently to your best friend and your bank, and so does everyone. Labelling the audience lets Claude learn the range, so it can ask “is this a mate-message or a client-message?” instead of averaging you into someone who sounds mildly odd to everybody.

Step 3: Build your personal ban list

What it is: the flip side of your voice - the words you’d never say. Every AI-written text is sprinkled with words that are fine in general but wrong for you, and a personal ban list catches them.

How to do it: straight after the mirror prompt, in the same chat:

Nick this one:

Based on those three messages, list 10 to 15 words and phrases that AI writing tends to use but that I clearly never would. Be specific to me, not generic - base every entry on evidence from how I actually write. This is my personal ban list: when you write as me, these never appear. Show me the list so I can veto anything I'd actually say.

The bit everyone gets wrong: copying someone else’s ban list off the internet. Don’t do it - their tells aren’t your tells. Maybe you genuinely do say “no worries at all” six times a day; banning it would make AI-you less like real-you. The whole point is that this list is grown from your own messages, which is why the prompt asks for evidence and why you get a veto. Your ban list should be as personal as your handwriting.

Step 4: Save it, then summon it

What it is: making the voice note permanent, so you never do this exercise again.

How to do it: once the “How I write” note and ban list look right, say: “Save this to your memory - use it whenever I ask you to write as me.” (Memory needs to be on: Settings → Capabilities, free for everyone. If you’d rather, paste the note into Settings → Profile instead, or into a Project’s instructions.) From then on, the trigger phrase is “in my voice”: “Reply to this in my voice.” “Draft the newsletter in my voice, client version.”

The bit everyone gets wrong: never testing it. Have Claude write something in your voice and show it to someone who knows you well, without saying what it is. If they don’t blink, you’ve done it. If they say “since when do you say ‘furthermore’?”… back to the ban list with the evidence.

A worked example

Say you run a dog-walking round. Your three messages might be: a WhatsApp to a client (“Bertie was an absolute mud missile today, hosed him down, he’s fluffy and shameless”), an email answering a new enquiry, and a longer message to the walker who covers your holidays. The note that comes back might say things like: opens with the dog’s name not a greeting, short sentences, jokes to soften bad news, never uses exclamation marks in emails but scatters them in texts, signs off differently per audience. That’s the stuff you could never have listed about yourself… and it’s exactly what makes writing sound like a person.

The small print

It won’t be perfectly you on day one - expect to say “I’d never phrase it like that” a few times in the first week, and each correction sharpens the note. In long conversations it can drift back towards formal, so a quick “in my voice, remember” resets it. And some messages shouldn’t be delegated at all: the genuinely emotional ones deserve your own slow typing.

The one move

Dig out the three messages. That’s the step people stall on, and it’s ten minutes of light nosiness through your own sent folder. Once they’re pasted, the rest is downhill.

One distinction before you go: this guide clones your voice. If what you’ve got is a lump of AI text that just needs to sound human in general, that’s a different fix - and it’s all in Stop AI Writing Like AI: The Plain-English Fix.