Writing Emails With AI That Still Sound Like You

A model has no idea how you write until you show it, and describing your own voice in adjectives does not work. Eight of your own emails does. The procedure takes twenty minutes once and is reusable indefinitely.

Why the default sounds like nobody

With no instruction about voice, output lands on the average of an enormous amount of written English, which for emails means corporate customer-service register: “I hope this message finds you well”, “I wanted to reach out”, “please do not hesitate”. It is not wrong. It is unmarked, and unmarked prose from somebody the recipient knows reads as either distant or outsourced.

Adjectives do not fix it. “Write in a friendly, professional, concise tone” is what everybody asks for, so it lands in the same place. Voice is carried by specifics — sentence length, whether you greet by name, whether you apologise, how you close — and those are far easier to extract from examples than to introspect about.

Building a voice sample from your sent folder

1. Pick eight emails you actually sent and were pleased with. Spread them: two to a close colleague, two to a customer or client, two delivering something awkward, two short logistical ones. Range matters more than volume.
2. Redact them. Replace names with PERSON_A, companies with COMPANY_A, and change or remove figures, addresses and anything confidential. Ten minutes with find-and-replace. This is what makes the rest of the procedure safe.
3. Keep them in one file on your own machine, each labelled with its context: [to a client, delivering bad news]. The label matters as much as the text, because it tells the model which sample to imitate for which situation.
4. Refresh it annually, or when you change job. Your register moves and a two-year-old sample starts producing a version of you that colleagues notice.

Eight is enough. The gain from more is small and the cost is that a long sample crowds out the actual request — a real constraint worth understanding once, in what the context window is and is not.

Turning the sample into rules

You can paste the samples every time, and for occasional use that is fine. Better is to extract a set of rules once, because rules are short, re-readable, and you can correct them.

Below are eight emails I wrote. Describe how I write, as a list of
rules another writer could follow. Be specific and mechanical.

Cover at least:
- typical sentence length, and whether it varies
- greeting: name only, name plus hello, nothing
- how I open: straight into the subject, or context first
- whether I use contractions, dashes, semicolons, exclamation marks
- how I say no, and how I apologise (or whether I do)
- typical email length in sentences
- how I close, and whether I sign my name
- three phrases I use that another writer might not
- anything I never do

Where the eight emails disagree, say so and note which context each
version belongs to.

[eight redacted emails, each labelled]

You will get twenty rules. Read them — this is a genuinely interesting exercise in itself — and delete the ones that describe habits you would rather drop. Keep the file. It is the only artefact you need.

A real example of the kind of output that is worth keeping, so you can judge whether yours is specific enough:

VOICE RULES
- Opens with the first name and a dash: "Ana - ". Never "Hi" or "Dear".
- First sentence is the point. No context paragraph before it.
- Sentences average 12 words. Frequent one-line paragraphs.
- Uses contractions everywhere except when declining something.
- Never uses exclamation marks. Never uses "just" as a softener.
- Says no directly, then gives one reason, then offers the nearest
  thing that is possible.
- Closes with "Thanks -" and the first name, or with nothing at all
  on short internal emails.
- Recurring phrases: "worth a word with", "not a problem this time",
  "before I chase anyone".
- Never apologises for delay unless the delay caused someone work.

A worked before and after

The task: telling a client that a deadline will slip by a week.

Without the voice rules, the request “write an email telling a client the project will be a week late” produces the familiar shape: an opening line hoping they are well, a paragraph of context, an apology, a passive construction that avoids saying who is late, a reassurance, and an offer to jump on a call.

With the rules and the facts:

Write an email using the VOICE RULES below.

Facts (use only these):
- The migration will finish 12 August, not 5 August.
- Reason: the client's finance system export was delivered on the
  3rd rather than the 28th, which was outside our control but I am
  not going to make a point of it.
- Nothing else in the project is affected.
- I want to confirm the 12th, not open a negotiation.
- Their contact is Ana; we speak weekly.

[voice rules]

What comes back is four sentences that open with the date, state the cause without blame, confirm rather than ask, and close the way you close. You change a word and send it. The difference between this and the first version is entirely in the inputs, which is the general lesson of writing a prompt that gets a useful answer.

Replies, which is where the time actually goes

Almost nobody spends their day composing emails from nothing. They spend it replying, and replies have a property that makes them the best case for this: the material is already in front of you. Paste the incoming email, your decision in five words, and the voice rules.

Reply to the email below, using the VOICE RULES.

My decision: no to the discount, yes to the extended payment terms,
do not mention the March order.

Rules:
- Answer every question they asked. If I have not told you the
  answer to one, write [ANSWER?] rather than guessing.
- Do not commit me to anything I have not stated above.
- Do not repeat back what they wrote.

[incoming email]
[voice rules]

The five-word decision is the whole method. It takes ten seconds and it is the part that only you can supply; everything else is transcription. The [ANSWER?] marker catches the specific failure of drafted replies, which is an invented answer to a question you had not noticed in paragraph four of a long email.

Two habits are worth adopting alongside it. Read the draft against the incoming email question by question before sending, because the commonest complaint about a reply is not its tone but that it answered two of three questions. And never let a drafting tool send anything directly: the moment a human stops reading the outbound message, every protection on this page stops applying.

Doing this without leaking your correspondence

Your sent folder is other people’s correspondence as much as your own, and it is the single most sensitive text most people own. Three rules make the procedure defensible.

• Redact before, not after. Once text has been sent to a provider, no setting retrieves it. Deleting a conversation deletes your view of it and does not necessarily reach backups, logs or anything already used for training under the terms you accepted. This distinction is the whole subject of using AI without handing over your data.
• Never sample from privileged or regulated correspondence. Legal advice, medical information, HR investigations, safeguarding. Sample from your ordinary logistics emails instead; voice is voice and it shows up just as clearly in an email about a delivery date.
• Check what your employer permits before pasting work email anywhere. Many organisations have a rule and most staff have not read it; a good internal policy answers this specifically, and what such a policy contains is worth knowing if you are the person who has to write one.

Emails you should never draft this way

Not because it would not work, but because the work is the point and the recipient can tell.

• Condolence, illness, bereavement. A short, awkward, obviously handwritten message is received well. A fluent one is received badly, and if it is ever recognised as generated the damage is permanent.
• A genuine apology. An apology is a statement about what you understand you did. Outsourcing the sentence outsources the understanding, and it reads as the corporate non-apology it structurally is.
• Disciplinary, dismissal and grievance correspondence. These have legal effect, they are read closely afterwards, and their exact wording matters. Draft from your organisation’s template and take advice.
• Anything you have not read. Obvious, universally violated. You are the sender. Every word is yours the moment it leaves.