How to Give AI Your Writing Style (So It Actually Sounds Like You)

The first time I asked an AI to write like me, it handed back a stranger wearing my jacket.

I'd pasted three blog posts into the chat and asked for a fourth "in my style." What came back had my vocabulary and none of my pulse. Every sentence ran the same length. The transitions were things I would never say. It opened with a thesis, which I never do, and closed with a summary, which I'd rather quit writing than do. It read like someone doing an impression of me at a party. Recognizable, and wrong.

I've been writing professionally for twenty years — newspapers, marketing, two novels, a travel blog that's older than my kid. So my first reaction was the writer's reaction: see, it can't do this. But that's not what happened. The AI didn't fail. My input did. I'd asked it to reverse-engineer a voice from three samples and a vibe.

How do you give AI your writing style?

You don't describe your voice. You document it. Collect a real body of your writing, have the AI extract the patterns you can't see yourself, and save the result as a writing codex: a plain file of rules and example passages that you own. Then paste that file into whatever tool you're using, today and next year. Capture the voice once. Reuse it everywhere.

That's the whole method. The rest of this post is how to do it well, and why the obvious shortcuts don't hold.

Why doesn't "write in my style" work on its own?

Two reasons, and the first one stings: you are the worst-placed person to describe your own voice. You've never heard your own accent. Ask most writers how they write and you'll get adjectives — "conversational," "direct," "a little dry" — which describe roughly every newsletter on the internet. The things that make you sound like you are mostly invisible to you: how your sentence lengths vary, which punctuation you lean on, the three transitions you reach for without noticing, the words you'd never be caught using.

The second reason is the one-sample problem. A single piece of writing shows the AI a topic, not a pattern. Style is what stays constant when the topic changes. Feed a model one essay and it learns what you wrote about. Feed it fifteen pieces across contexts and it can start to see how you write about anything. I made the full argument in how to make AI sound like you. The short version: voice lives in patterns, and patterns need a corpus.

Step one: collect the evidence

Ten to twenty pieces, and the spread matters more than the count. Pull emails you're proud of, posts that got the reaction you wanted, something long, something dashed off. Different registers, different audiences. You're building the dataset that lets the constant emerge. And the constant is the voice.

Skip anything heavily edited by someone else, and anything an AI helped write. You want the unfiltered signal.

Step two: extraction beats description

Now flip the job. Instead of telling the AI what your voice is, paste the samples in and make it tell you. Ask for the mechanics: average sentence length and how much it varies, punctuation habits, how paragraphs break, signature transitions, how the pieces open and close, where the humor sits, which words repeat. Then the part most people skip — ask it what you never do. The never-list is half the voice.

And then let it interview you. A good model asking pointed questions — "you use fragments for emphasis; deliberate or habit?" — will surface rules you didn't know you had. That guided excavation is context mining, and it works for voice exactly the way it works for the rest of your context: you don't find the patterns by staring at the blank page. You find them by being asked.

Step three: rules plus receipts

The output of the extraction becomes your codex, and the format matters. Rules alone aren't enough, because rules are interpretable. "Vary sentence length" can produce a hundred different writers. So the codex gets two halves: the rules, and the receipts. The receipts are five to ten real passages of your writing, labeled by context, that the AI can calibrate against.

Mine includes things like: periods as percussion. Oxford comma, non-negotiable. Semicolons banned as rhythmically neutral. Scene before thesis, always. And then the passages that show what those rules sound like when they're working. When a draft comes back off-key, the fix is usually not a new instruction. It's pointing at a passage and saying more like this.

Step four: test blind, log the misses

Load the codex into a fresh chat and have it draft something real. Read it out loud. You'll wince somewhere: a transition you'd never use, an ending that ties a bow on it. Good. Every wince is data. Add the correction to the codex, not just the chat, and run it again. Three or four rounds of this and the drafts stop being impressions and start being usable first passes. The codex becomes a scar log in the best sense: every miss it absorbs is a miss that never comes back.

Where should your voice actually live?

Not in any one tool. And this is the part that decides whether the work pays off once or forever. Rebuild your style in ChatGPT's custom instructions, again in Claude's preferences, again in whatever launches next spring, and you now have three drifting copies locked in three apps. I've compared the ways to give AI context, and the pattern holds everywhere: settings die with the tool; files travel.

So the codex lives as markdown you own. One file, one source of truth, pasted wherever you're working. It's one of the six context anchors for a reason — voice is the context people feel first. At RUMO, the Writing Codex anchor is the guided version of everything above: the interview, the extraction, the rules-plus-receipts structure, done in about half an hour instead of a Saturday. And it sits beside a personal constitution, because a tool that knows how you sound and what you value is when the drafts start getting eerie.

These days, when a draft comes back, I still read it out loud. Recently one made me pause for a different reason: I couldn't find the seam. Not my best writing — my Tuesday writing, which is exactly the bar. The stranger gave back the jacket.

Your voice took years to build. Documenting it takes an afternoon. Cheap rent for getting every AI you'll ever use to stop doing impressions.