Make AI Write Like You, Not Like Everyone Else

You feed the model a few notes, ask for a polished version, and what comes back is competent and completely anonymous. The rhythm is even, the vocabulary sits in the safe middle of the road, and every other person using the same tool is getting writing that sounds exactly like it. For a freelancer or a team of one, that is a real problem, because your voice is part of what people are paying for. When your drafts stop sounding like you, the work stops feeling like yours, and clients can sense the difference even when they can’t name it.

This happens for a dull reason. A language model writes toward the average of everything it has read, so left alone it produces the most generic version of any sentence. Telling it to be “more casual” or “more professional” barely helps, because those words mean a hundred different things and the model fills the gap with its own default. The fix is not a better adjective. It is to hand the model a short, concrete rulebook pulled from your own writing, so it has something specific to copy.

That rulebook is your voiceprint. It takes about half an hour to build the first time, and after that you save it so it loads on its own. The steps are below.

Start from three things you actually wrote

Open three pieces of your own writing where you sounded like yourself. A couple of sent emails and a post you were happy with will do. Skip anything you wrote to sound formal for its own sake, and skip anything an editor heavily rewrote. You want your natural voice, not your committee voice. Three short samples are plenty, because you are looking for patterns, not volume.

Pull the rules out of your own writing

Now read those samples like a stranger and answer a few specific questions about how you write. How long are your sentences, and do you mix long ones with short, or keep them even? Which words show up again and again, and which ordinary words would you never use? How do you open a message before getting to the point? Do you use contractions, and how often? Do you hedge with “probably” and “usually”, or state things flatly?

Write the answers down as plain rules. Something like: “Sentences mostly short, one longer one per paragraph for rhythm. Always uses contractions. Never says ‘leverage’ or ‘utilize’. Opens with the point, not a greeting. Dry, understated jokes, no exclamation marks.” Six to ten lines like that is a working voiceprint. The reason this beats a tone word is simple.

Tone words are wishes. Rules are something the model can actually follow.

Save it where it loads every time

A voiceprint you have to paste in every time is a voiceprint you will quietly stop using. Both major tools let you save it once. In ChatGPT, drop it into Custom Instructions under your personalization settings, where it applies to every new chat and holds up to 1,500 characters. In Claude, the same thing lives under Instructions for Claude in your profile, or inside a single Project if you only want that voice for one client’s work. The two tools even start from slightly different house styles, which we compared in our honest look at Claude versus ChatGPT for writing. If the mechanics of writing the instruction itself feel shaky, the basics of a clear prompt are worth ten minutes first.

Do a cleanup pass on the obvious tells

Even with a good voiceprint, a few machine habits slip through, and they are exactly the things readers half-consciously register as “this was written by AI”. Strip the em-dashes it scatters around down to one a paragraph at most. Cut the throat-clearing openers that warm up for a paragraph before saying anything, and the three-part flourishes like “clear, concise, and compelling”. Delete the “this isn’t just X, it’s Y” construction every time you see it. None of this takes long once you know the patterns, and it is the same edit that separates writing that merely passes from writing that sounds like a person wrote it. The people who get the most out of these tools treat them like a colleague they are training, correcting the same tics until the model learns to stop making them.

A voiceprint will not make the model write better than you. It will stop it from writing blander than you, which for anyone with their name on the work is the part that matters. Build it once, save it, and the flat, anyone-could-have-written-this draft stops being the thing you start every job from.