How to Make AI Replies Sound Like You

AI drafts sound generic until you tell the tool who you are. How to describe your voice, set tone and length, and fix a draft in ten seconds.

· 4 min read · Writing

The usual complaint about AI-written replies is that they all sound the same: polite, complete, a little eager, and written by nobody in particular. That is a fair description of what a language model does when it knows nothing about you. It writes the average.

The fix is not a cleverer model. It is giving the model something to aim at, then spending a few seconds on the draft. Here is how to do both.

Why drafts sound generic

A model has read a great deal of writing and, asked for a reply, returns something near the middle of all of it. The middle is grammatical, agreeable and forgettable.

Your own writing is not in the middle. You have subjects you return to, words you like, things you would never say and a usual length. Each of those is a piece of information the model does not have until you supply it.

Describe your voice in four parts

Most tools give you a box for a persona or a style. People leave it empty, or write "friendly and professional", which describes everybody. A useful description has four parts.

1. What you do and who you talk to.

I build developer tools at a small startup. I mostly talk to other engineers and a few founders.

2. How you sound.

Dry, direct, a bit skeptical. Short sentences. I joke, but I do not do enthusiasm.

3. What you bring.

I usually answer from experience: what we tried, what broke, what it cost.

4. What you never do.

No exclamation marks. No emojis. No hashtags. I never say "game changer" or "great post".

Four or five lines like these change the drafts more than any other setting. The last part matters most. Telling a model what to avoid is the quickest way to take the sameness out.

Set your defaults once

Voice is also shape. Decide these once so that you are not fixing them on every draft:

  • Length. If you write one-liners, set replies to short. A three-sentence draft from a one-line person looks wrong at a glance.
  • Tone. Pick the one you use most as the default, and switch only when a post calls for it.
  • Emoji. Never, sometimes or often. Most people who say "sometimes" mean "rarely".
  • Language. If you reply in the language of the post, turn that on.

Tell it your habits

Small habits are what friends would recognize you by. Put them in the description:

  • You write in lowercase.
  • You start with the point and skip the greeting.
  • You use "we" for your company.
  • You ask a question more often than you make a statement.
  • You use numbers where you can.

None of these is deep. Together they are a voice.

The ten-second edit

Even a good draft improves with a quick pass. You do not need to rewrite it. Do one or two of these:

  1. Swap one word for yours. If the draft says "significant" and you would say "huge", change it.
  2. Cut the first clause. Drafts often open with a warm-up such as "That is a fair point, but". Start at "but".
  3. Add the thing only you know. A number, a name, what happened when you tried it. This is the part no model can write.
  4. Break a perfect sentence. People on X write fragments. One is fine.
  5. Delete the last sentence. If the reply still works, it was not needed.

Use the other drafts

If a draft is close but not right, ask for another before you start editing. In XAICG, one request brings back several drafts and stepping through them is free, so a second look costs nothing. Often the second or third has the angle you wanted.

If none of them fit, that tells you something. Either the post is not one you have much to say about, or your description needs another line.

When to write it yourself

A few kinds of reply should come from you from the first word:

  • Condolences and apologies. People can tell, and they should not have to wonder.
  • Claims about your own work. Only you know what is true.
  • Anything that could be read as advice on health, law or money. Your name is on it.
  • A reply to a friend. The draft will be fine, and it will not be you.

A worked example

The post and both replies below were written for this article, to show the difference a description makes.

The post:

Unpopular opinion: daily standups are a status report for managers, not a tool for the team.

A draft with no description of the writer:

Great point! Standups can definitely become status reports if not managed well. It's important to keep them focused on collaboration.

The same post, with the four-part description above:

ours turned into a status report the week we passed 8 people. killed it, kept a written update. nobody has asked for it back.

The second one has a number, an event and an outcome, in the writer's own shape. Most of that came from the description. The number came from the writer, in the ten-second edit.

That is the whole method: tell the tool who you are, set your defaults, and add the one thing you know. For the patterns to watch for in any draft, see signs a reply was written by AI.