How to Get ChatGPT or Claude to Consistently Write Like You

Sounding less robotic and sounding like a specific person are two different problems. Most people only ever solve the first one, then wonder why the output still doesn't feel like theirs.

Ask a model to "write like me" and it has almost nothing to work with. It doesn't know what that means, and to be fair, most people couldn't fully articulate it either if you asked them directly. Personal voice is made of dozens of small habits nobody consciously tracks: how long your sentences usually run, which words you reach for without thinking, whether you soften a strong opinion or state it flat, what you do at the start of a paragraph versus the end. None of that shows up in an instruction like "casual but smart." It shows up in your actual writing, which is where any real fix has to start, and it takes a little deliberate effort to notice in the first place.

Why this is harder than sounding less robotic

Making AI writing sound more natural in general is a solved, well documented problem: vary sentence length, cut the hedge phrases, add specific detail. That gets you from robotic to competently human. It doesn't get you to a specific human, because "natural" has a huge range and your particular spot in that range is not the same as anyone else's. Two people can both write naturally and sound nothing alike. The gap between generic-natural and recognizably-you is the part most advice skips entirely.

Finding your own patterns first

Before a model can imitate your voice, you have to actually notice what it is, which is harder than it sounds since most of it is invisible to you by default. Pull together five to ten pieces of your own writing that you're actually happy with, emails, posts, notes, whatever exists, and read them looking specifically for pattern, not content.

• Sentence rhythm. Some people write short and punchy by default, others long and winding with several clauses stacked up. Most people are more consistent here than they'd guess.
• Openers and closers. How do your paragraphs usually start. How do they usually end. A lot of personal voice lives specifically in these two spots.
• Asides and parentheticals. Do you interrupt yourself, qualify things in the middle of a sentence, add a quick aside before continuing. This is one of the most distinctive and most commonly overlooked habits.
• Words you overuse. Everyone has a small personal vocabulary of go-to words and phrases, usually invisible to the person using them and obvious to anyone else reading a stack of their writing side by side.
• Where you're direct versus where you hedge. Some people state opinions flatly and hedge on facts. Others do the opposite. This is a real, consistent pattern for most writers, not a random mix.

Turning that into an instruction that actually works

Once you've noticed the pattern, the instruction that captures it is almost never a list of adjectives, its the actual samples plus a short, specific note about what to notice in them. "Match this voice" followed by three real paragraphs does more than a paragraph of self-description, because the samples carry information a description can't: the exact rhythm, the exact vocabulary, the exact structure, all at once, without you having to correctly name any of it first.

A useful format: paste two or three short samples, then add one or two sentences pointing at the specific things you want matched, not described in general terms but named directly, "notice the one-line opening paragraph" or "notice how the second-to-last sentence usually undercuts the one before it." Naming the specific pattern out loud, even briefly, seems to help a model weight it more heavily than leaving it for the model to infer entirely on its own from the samples.

Making it stick across sessions

A voice profile that only works inside one conversation isn't very useful if you're writing regularly. Most major AI tools now offer some way to save a standing instruction that carries into new conversations rather than something you have to re-paste every time, Claude's user preferences and Projects are one version of this, and other assistants have their own equivalents under different names. Whatever the specific feature is called, the mechanic is the same: write your voice profile once, save it somewhere persistent, and it applies automatically going forward instead of depending on you remembering to paste it in.

Worth being deliberate about what goes into that saved version. A profile built entirely from one email you happened to have handy will overfit to that email's specific mood and context. Pull from writing across a few different situations, casual and slightly more formal, if both exist in your own work, so the saved profile captures the actual range rather than one narrow slice of it.

Testing whether it actually worked

The honest test isn't whether the output sounds good in isolation. Plenty of generic AI writing sounds perfectly fine on its own. The test is whether it sounds like a specific person, which only shows up in comparison. Put a real sample of your writing next to the model's output on the same rough topic and read both. The differences that jump out, a phrase you'd never actually use, a sentence structure that's too neat compared to how you really write, a transition that's smoother than yours usually are, are exactly the signal you need to go back and adjust the profile.

This comparison step is worth doing more than once, and worth doing with different kinds of writing, a quick note, a longer explanation, something slightly more formal, since a profile that nails one register can still miss badly on another. Treat the first version of a voice profile as a draft in its own right, not a finished product, the same way you'd treat any other first attempt at describing something as fuzzy and personal as your own writing style.

Common ways this goes wrong

• Overfitting to one register. A profile built from casual texts will produce oddly casual formal emails. Feed it a range if your own writing has a range.
• Copying surface quirks without the structure underneath. A model that's told to use your favorite word a lot but isn't shown your actual sentence structure ends up sounding like an impression of you rather than your writing, technically similar on the surface and wrong underneath.
• Treating it as finished after one attempt. A voice profile is worth revisiting the same way any instruction set is. If output starts drifting or never quite landed in the first place, update the samples rather than assuming the first attempt was as good as it gets.
• Never actually testing it blind. Show the output to someone who knows your writing without telling them it was AI-assisted. If they can't tell, the profile is working. If they immediately flag something as off, that's specific, useful feedback about what the samples missed.
• Assuming one profile covers everything you write. A voice built for casual notes and a voice built for careful analysis are not always the same voice, even for the same person. If you write in noticeably different registers depending on context, its worth keeping more than one saved profile rather than forcing a single version to stretch across all of it.

The actual goal

Getting a model to sound like you isn't really about hiding that AI was involved. The actual point is not having to rewrite every single draft by hand just to make it feel like something you'd actually publish. The work is mostly upfront: noticing your own patterns, saying them out loud instead of trusting them to be obvious, and giving the model real samples instead of a vague description. Once that's done and saved somewhere persistent, it holds up remarkably well across a lot of ordinary, everyday writing, and it stops being something you have to think about every single time you open a new conversation.

Writing for a team rather than just yourself runs into a related but different problem, one person's voice profile doesn't automatically transfer to a group. Here's the version of this for keeping a brand voice consistent across multiple writers.