AI doesn't know your voice.
That's not a flaw. It's a starting condition. Claude and ChatGPT default to what they learned from billions of web pages, which is "professional internet writing," polished, slightly eager, relentlessly cheerful, with the same rhythm every other AI-generated post uses. The voice that makes your readers stop scrolling lives in the gap between your actual writing and that default.
Most people try to close the gap in every single prompt. They add "match my style" to the end. They paste a few sentences as an example. They reject the first draft and fight for 20 minutes to drag it closer to something they'd publish. Then they do it again tomorrow.
The shortcut nobody runs is to spend 20 minutes one time, extract your actual voice patterns into a written guide, and deploy that guide everywhere you use AI. After that, every output starts inside your voice instead of trying to fight its way in.
Here's the workflow.
Why describing your style yourself doesn't work
If asked, you'd probably describe your writing as "casual, direct, with some humor." That description is true. It's also useless to AI, because every internet writer would describe themselves that way.
Your real voice is more specific. Your average sentence length. The words you reach for and the ones you never touch. How you open paragraphs. How you handle uncertainty. The structures you repeat without realizing it. The things you never do.
You can't tell AI these things from memory because most of them are unconscious. You wrote that way for years without noticing the pattern. The fix is to have AI extract the patterns by looking at samples of your work, which is the thing AI is actually good at.
Step 1: Gather 5 samples (5 minutes)
Pull together five pieces of writing that you'd be happy to be imitated by. Articles, LinkedIn posts, newsletters, work documents that sound like you, even long messages you've sent. They should all be over 200 words and ideally cover similar ground (all professional writing, or all casual writing, but not a mix).
Two things matter most. They should sound like you on a good day, not your average day. And they should be in the register you actually want AI to write in. Don't include a wedding speech if you'll never have AI write wedding speeches.
If you have access to your past AI chat history, there's a sneaky alternative worth knowing about. A 2026 writer named JuanjoFuchs realized that his dictated chat messages were a more honest sample of his actual speech patterns than his polished articles. He exported 1,012 Claude prompts and mined them for patterns. The dictated stuff was closer to his real voice than the carefully edited published stuff.
For most people, five polished samples is the right starting point. Save the chat-mining trick for round two.
Step 2: The extraction prompt (10 minutes)
This is the key step, and the one that earns the whole exercise.
"You are a voice analysis expert. I'm sharing five writing samples I produced. Extract my voice profile across these six dimensions:
- How I open paragraphs. What's the first sentence usually doing?
- Sentence length patterns. Average, range, rhythm.
- Moves I repeat. Do I enumerate? Use parentheticals? Quantify?
- Vocabulary I reach for vs vocabulary I avoid.
- How I handle uncertainty and evidence.
- What I consistently never do.
For every observation, quote the actual phrase or pattern from my samples as evidence. 'Uses short sentences' is useless. 'Opening sentences average 8 words; elaboration sentences average 18' is useful. If you can't ground an observation in a real quote, leave it out.
Format as six headers, two to four observations per header."
The "quote my text as evidence" line is the entire point. Without it, AI gives you generic style advice that could describe anyone. With it, AI is forced to look at your actual patterns.
Paste your five samples. Run the prompt. You'll get back a structured analysis. Read it carefully. Some observations will feel obvious. Some will feel surprising, like AI noticed something you'd been doing for years without naming. The surprising ones are usually the most valuable.
This output, plus a short "ban list" of words and phrases you don't want AI to use, is your style guide.
Step 3: Refine through testing (5 minutes)
The style guide isn't done at first generation. You test it.
Paste the extracted profile back into a fresh chat. Ask the AI to write something you've already written, like a paragraph from a recent piece. Compare the new output to your original.
Where does it miss? Maybe it's too formal. Maybe the rhythm is wrong. Maybe it's missing your habit of opening with a one-line punch. Wherever it misses, add a rule to the profile that addresses the gap.
Sophia Cave, a 2026 writer at Like One, put this directly. Three rounds of test-and-update and your profile is production-ready. The third round usually catches the subtle stuff that the first two missed.
Most people skip this step. The result is a style guide that looks complete but produces work that's almost-but-not-quite-you. The 5 minutes you spend on refinement is the difference between a guide that works and one that almost works.
Step 4: Deploy it (where it lives forever)
The style guide is a document. It needs a home.
For Claude users, the easiest path is to drop the guide into your Project instructions, or save it as a Claude Skill called "Brand Voice." Every chat in that Project (or whenever the Skill auto-triggers on writing tasks) inherits your voice.
For ChatGPT users, paste the guide into Custom Instructions, or into the Instructions field of a Custom GPT or ChatGPT Project. Same effect, slightly different button location.
For Gemini users, paste it into a Gemini Gem.
For one-off tasks in other tools, or if you're using AI inside a third-party app, keep the guide in a Notion page or a sticky note and paste it at the top of any new prompt.
The deployment is the part most people skip after building the guide. Don't. The guide is useless if it lives in a doc you never paste anywhere. The point is to make every AI session start inside your voice by default.
The quarterly refresh
Your voice evolves. The phrasing you used heavily six months ago might feel stale now. New patterns might have crept in. Old rules might no longer apply.
Every three months, rerun the extraction prompt on five fresh samples. Compare the new profile to your old one. Update the rules that drifted. Cut the rules that no longer apply.
This takes 15 minutes a quarter. It keeps the style guide aligned with how you actually write now, not how you wrote a year ago.
The mistake that wastes the setup
Building a 50-rule style guide.
If your profile has dozens of rules, AI starts tripping over them. It tries to obey rule 7 while violating rule 12. The output gets stiff. The voice gets lost in the rule-following.
A working guide is usually 8 to 15 rules across the six dimensions. Tight, specific, evidence-backed. If the guide is longer than a single page, you're describing a style, not encoding one.
Less is more. Quote your evidence. Refresh quarterly. Skip the rest.
If you want the broader workflow of using AI for daily and weekly habits (including the Personal Context Profile that contains your voice rules alongside who you are, what you do, and what you're working on), I just published Your AI Operating System: The Beginner's Field Guide to Letting AI Do Your Busywork on Gumroad. Volume 1 of my AI for Real Life library. Volume 2 will focus specifically on AI for content and writing, where this style guide workflow becomes one of the central tools. Launch price for Volume 1 is $19, with every future volume in the library free for existing buyers.
AI tone is the battle most people lose every day.
Twenty minutes once. A style guide deployed everywhere. The fight is over.
Tags: Artificial Intelligence, Writing, AI Tools, Productivity, Workflows