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

Ask Claude to write an email and it reaches for its poshest register. Out comes something you’d never say to a client you actually like.

Describing your voice doesn’t fix it. Try describing your own accent and see how far you get. Showing it the evidence does: thousands of words you’ve already written, at normal speed, without once thinking about “tone”.

This guide builds a written description of how you write, once, that you save and reuse forever. If what you’ve got instead is one draft on screen right now that reads like a leaflet, that’s a different job and it’s over here: Stop AI Writing Like AI.

Step 1: Dig out three messages (five minutes)

Open your sent folder and your messages, and copy out:

1. One quick one. A WhatsApp or text, a few lines, to someone you’re relaxed with.
2. One everyday email. To a client, a supplier, the school office. Something routine.
3. One longer one. Where you explained something, or handled something delicate.

Pick your normal writing, not your best. Feed it your fanciest work and you’ll get an AI that impersonates you showing off, which is its own kind of wrong. The version of you people recognise is Tuesday-afternoon you, banging out a reply between two jobs.

Before you paste, swap out anything you wouldn’t want on a sticky note: surnames, account numbers, the rate you quoted someone. Your voice survives the redaction fine.

Step 2: Run the mirror prompt

Nick this one, fill in the blanks:

You're studying how I write, so that later you can draft as me and nobody blinks. Ignore what these messages are about. Look only at how they're built.

Message 1, a quick one to a friend: [PASTE: e.g. the WhatsApp where you moved Saturday]
Message 2, an everyday email to a client: [PASTE: the one confirming a job, sent in a rush]
Message 3, a longer one where I explained or handled something: [PASTE]

Give me:
1. How I open and how I close, with a quoted example of each.
2. My sentence rhythm: typical length, where I run long, where I go short.
3. Words and phrases I reuse, quoted exactly.
4. How I soften bad news, or whether I don't bother.
5. My punctuation habits, including what I never use.
6. How my voice shifts between the three audiences, one line each.
7. Then write it up as a note called "How I write", short enough to paste into a settings box, with the evidence quotes left in.
8. Give me that note as a document I can download and keep.

Rules: quote my actual words as evidence for every single observation, and never tidy a quote up. English spelling. Don't tell me my writing is lovely, I need accuracy, not encouragement. If something only shows up in one of the three messages, say so instead of calling it a rule.

Ask me up to three questions before you start.

The bit everyone gets wrong: dropping the “to a friend” and “to a client” labels. You write differently to your mate and to your bank. Label the audience and Claude learns your range, so it asks “client version or mate version?” instead of averaging you into someone who sounds slightly odd to everybody.

Step 3: Build your ban list

Same chat, straight after.

Then this one:

You're building my personal ban list: words that turn up in AI writing that I plainly never use. Work only from the three messages and the note you just wrote.

Give me:
1. Ten to fifteen words or phrases to ban, each with one line on why it clashes with how I actually write.
2. What I say instead, lifted from my messages where possible. If "reach out" is on the list and I say "give us a shout", put those next to each other.
3. Five words I DO use that a stripped-back AI voice would wrongly delete. These are mine, protect them.
4. Then give me the whole thing as five lines I can paste onto the bottom of my "How I write" note.

Rules: specific to me, never a generic list off the internet. If you can't point at evidence, leave it out. Show me the list before treating it as final, so I can veto anything I genuinely say.

Ask me anything you need to know first.

Why point 3 is in there: ban lists get overzealous. Most of the ones on the internet would strip “no worries” out of somebody who says it eight times a day, and banning your own habits makes AI-you less like you.

Step 4: Save it, then summon it

1. Say: “Save this to your memory, and use it whenever I ask you to write as me.” Memory lives in Settings, then the memory option.
2. Belt and braces: paste the note into Settings → Profile too, or into a Project’s instructions.
3. From then on the trigger is three words. “In my voice.” Reply to this in my voice. Draft the newsletter in my voice, client version.

Then test it. Have Claude write something short as you and show it to someone who knows you, without saying what it is. No reaction means it worked. “Since when do you say furthermore?” means back to the ban list, with that sentence as evidence.

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 week one, and hand each of those back with the correction, because that’s what sharpens the note.

In long conversations it drifts back towards formal. “In my voice, remember” pulls it straight.

And some messages shouldn’t be handed over at all. The genuinely emotional ones deserve your own slow typing.

Tonight

Go and find the three messages. That’s the step people stall on, and it’s ten minutes of light nosiness through your own sent folder. The rest is downhill.

That’s Claude sounding like you in one chat. The Claude Kickstart sets your voice up properly and then puts it to work on five real jobs, with a playbook for your actual trade. £47, one weekend.