How to Train ChatGPT to Write Like You (and Where It Breaks)

By Jack Stovell · 2026-08-25 · Guides

You can absolutely get ChatGPT to sound more like you. Twenty minutes, no special tools, genuinely fine results for a lot of everyday writing. Here's the method, honestly, and then here's exactly where it stops working, because that second half is the one nobody tells you.

The DIY method, step by step

First, collect three to five samples of your own writing. Emails, blog posts, whatever exists already. Don't overthink the selection; just grab things that sound like you on a normal day, not your best-ever paragraph.

Second, write a style brief. This is the bit people skip and shouldn't. Describe your tone in plain adjectives: witty, warm, direct, whatever fits. Note habits too. Do you use contractions constantly? Short sentences? Do you open with a question half the time? Write it down like you're briefing a new hire who's never read a word you've written.

Third, paste both the samples and the brief into ChatGPT's custom instructions, or into a Project if you're using one. Give it the brief, give it the samples, ask it to write something new in that voice.

Fourth, iterate. First output won't be perfect. Tell it what's off ("too formal", "cut the exclamation marks", "shorter sentences") and regenerate. Most people land somewhere usable after two or three rounds.

That's it. That's the whole method. For a casual LinkedIn post or a quick email draft, this works fine. Say so plainly: if the stakes are low and the output's getting a human edit anyway, don't overcomplicate it.

So why bother reading the rest of this?

Because the method has a structural flaw, not a using-it-wrong flaw. And once you see it, you can't unsee it.

Where it actually breaks

Here's the thing: the model reads your samples once, at the start of the conversation, and then it's improvising from memory for everything after. There's no constraint sitting underneath each new sentence checking "does this match?" It's pattern-matching off a vague impression, not measuring against your actual habits. That's why voice drifts. Not might drift, does drift, reliably, the longer the output runs or the more sessions you use it across.

Ask for 2,000 words and watch what happens by paragraph twelve. The contractions creep back in (or out). Sentence length creeps toward the model's default rhythm, which is smoother and more balanced than most real writers actually are. By the end it's writing like ChatGPT with your vocabulary sprinkled on top, not writing like you.

Cross-session drift is worse. Custom instructions get referenced loosely, not enforced. Ask it to write something Tuesday and something Thursday and compare them. Often they don't sound like siblings.

Then there's the adjective problem. "Witty but warm" means something different every single time the model interprets it, because adjectives aren't measurements, they're vibes. One generation reads witty as sarcastic, another reads it as playful, another decides warm means padding every sentence with reassurance. You didn't change your brief. The interpretation moved anyway.

And here's the practical kicker: custom instructions have a length limit. So you're forced to compress "how I actually write" (which might genuinely take pages to describe properly) down into a handful of adjectives and a couple of sample paragraphs. That's not a minor inconvenience. That's the whole problem in miniature: you can't fit stylometric reality into a text box built for short prompts.

So the honest version of this article's argument is: the DIY method fails not because you did it badly, but because "read some examples and hold a vibe" was never going to survive 2,000 words or three separate Tuesdays.

The alternative: measure it instead of describing it

This is what I build, so take it as the obviously biased comparison it is, but the logic still holds regardless of who's saying it. Instead of compressing your voice into adjectives, you can measure it. Sentence length variance, contraction rate, how often you use fragments, comma density, how you open paragraphs, that sort of thing, extracted directly from real samples of your writing rather than guessed at.

ScriptGrain does this by pulling 45 explicit attributes out of your writing through a two-pass stylometric analysis, then enforcing those as hard constraints on every generation, not just the first one. The difference isn't "better vibes". It's that the model has actual numbers to check against instead of a paragraph of adjectives it's free to reinterpret.

You don't have to take that on faith, either. There's a free browser-based tool at scriptgrain.com/tools/writing-style-analysis that measures 12 of those attributes client-side, no signup, nothing sent anywhere. Paste some writing in, see the numbers. It won't tell you everything, but it'll show you the measurable half of what "witty but warm" was always trying, badly, to describe. Beyond that, the free tier gives you one full profile and one analysis, no card required, if you want to see what the other 33 attributes look like.

Fair enough, you might say, but do I need this?

For a casual post, no. Honestly, no. The DIY method will do the job and you shouldn't feel bad about using it. But for anything long-form, anything published under your name repeatedly, anything where "sounds like me" actually matters to your reputation, description starts to buckle exactly where measurement doesn't. That's the whole distinction, really. Adjectives are opinions. Attributes are data. And at the end of the day, only one of those holds steady past paragraph twelve.