How to make Claude write in your voice

What detectors actually do

Two things are true at the same time here, and holding both is what keeps you sane.

They do get real human writing wrong, and unevenly. Stanford researchers ran seven detectors across essays written by non-native English students and found the tools "classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated", with 97% of those essays flagged by at least one detector. Vanderbilt turned Turnitin's AI detector off over exactly this, noting that even at the vendor's own claimed rate "around 750 student papers could have been incorrectly labeled as having some of it written by AI". OpenAI retired its own detector entirely, saying it was "no longer available due to its low rate of accuracy". So if your genuine work got flagged, that is a real thing that happens to real people, and it is not evidence about you.

And the arms race is not one you can win from the outside. Turnitin publishes its own numbers: a document-level false positive rate of "less than 1% for documents with 20% or more AI writing" and a sentence-level rate of "around 4%". More to the point, in 2024 they shipped a feature aimed squarely at humanizer output, built to identify "instances where AI paraphrasing may have been used to modify AI-generated text in submitted content". That is the whole business model of those monthly subscription sites, and it is being actively targeted by a company with far more resources than they have.

Pick three samples, and pick them well

The whole system is only as good as what you feed it. Three pieces is the minimum that lets Claude tell a habit apart from a one-off, and the pieces matter more than the number.

1. 📝 One thing you wrote carefully. An essay, a report, a long post, a cover letter you were proud of. This shows your structured voice: how you build an argument and how formal you get when you are trying.
2. 💬 One thing you wrote fast. A long message to a friend or colleague, a Slack explanation, a comment you left somewhere. This is where your real rhythm lives, and it is the part that makes the output sound like a person.
3. 🎯 One thing in the register you want. If the goal is better emails, use your best email. If it is essays, use your best essay. The style guide will lean towards whatever you give it most of.
4. 🚫 Nothing you did not write. Sounds obvious, and it is the mistake people make. Feeding it something AI helped with teaches it to imitate the thing you are trying to escape.

Have Claude build your style guide

This is the prompt that does the real work. It is long on purpose: a vague request gets you a vague guide that says you write "clearly and engagingly", which describes nobody.

I am going to paste three things I wrote myself. I want you to reverse-engineer how I write and turn it into a style guide you can follow from now on.

Analyse them properly rather than complimenting me. Measure my average sentence length and how much it varies, because that variation is the most distinctive thing about anyone's writing. Note how I open and close pieces, whether I use questions, how often I use contractions, my punctuation habits including whether I use semicolons or dashes at all, and how formal I get. Pull out the words and phrases I actually reach for repeatedly, and separately note the ones I clearly avoid. Tell me what I do when I am explaining something difficult, and what I do when I am making a point I care about.

Then write the style guide as direct instructions to you, in the second person, roughly 300 to 500 words, with concrete rules rather than adjectives. "Keep sentences short" is useless. "Average 14 words, but break to 4 words when landing a point" is usable. Include a short do-not list of things that would immediately read as not me. Finish by quoting three lines from my samples that you think are the most characteristically me, and say why in one sentence each.

Here are the samples:
[paste sample 1]
---
[paste sample 2]
---
[paste sample 3]

Pressure-test it before you trust it

The first style guide is always slightly wrong. It over-indexes on whichever sample was longest, or it mistakes something you did once for something you always do. Ten minutes of correction here is what separates a guide that works from one you quietly stop using.

Now test that style guide. Using it, write me three short pieces on the same topic: a 150-word explanation of [something you know well], a short message declining an invitation, and an opening paragraph for something longer.

Do not tell me they are in my voice. Instead, for each one, point out the two specific choices you made because of the style guide, and the one place you were unsure whether I would write it that way. Then, honestly, tell me which parts of the guide were too vague to act on.

I am going to mark up the drafts and tell you what is wrong, and I want you to update the style guide from my corrections rather than only fixing the drafts. Ask me for anything you need to make a rule sharper, especially about tone in situations the samples did not cover, like disagreeing with someone or delivering bad news.

Save it so it applies to everything

A style guide sitting in one chat is useless by Thursday. There are two places to put it so it applies automatically, and doing both is the setup that actually sticks.

1. 📁 Create a project. In Claude, create a new project and call it something you will recognise, like "My writing". Projects work on every plan, including Free, where you can have up to five, per Anthropic's projects documentation.
2. 🧭 Paste the style guide into the project instructions. This is the field that gets applied to every conversation inside the project, so the rules are loaded before you type a word.
3. 📎 Upload your three samples to project knowledge as well. The instructions tell Claude the rules, the samples let it hear the music. Having both is noticeably better than either alone.
4. ✍️ Start every writing task inside that project. The habit is the hard part. If you draft in a normal chat out of muscle memory, you get chatbot voice back and wonder why the setup did not work.

The two prompts you will use forever

Once the project exists, almost everything you do is one of these two moves: write something new in your voice, or fix something that does not sound like you yet.

Draft [what you need: the email, the essay section, the post] using the style guide in this project's instructions. The context is [who it is for, what they already know, what you want them to do next].

Before you write, tell me in two lines what angle you are taking and what you are deliberately leaving out, so I can redirect you before you commit a thousand words to the wrong idea. Then write it, and afterwards flag any sentence where you had to guess at something factual, because I would rather fill a gap than discover an invented detail later.

Keep it in my register, do not raise the formality to be safe, and do not add a summarising final paragraph unless my samples show that I do that.

And the one that quietly does the most work, on anything you or anyone else drafted:

Here is a draft. Compare it line by line against my style guide and tell me where it does not sound like me: [paste the draft].

Give me a table with three columns: the line as written, what is wrong with it in my voice specifically, and a rewrite. Only flag lines that genuinely break a rule in the guide, because if you flag everything I will stop reading. Pay particular attention to sentences that are all the same length, openings that state what the piece is about instead of starting it, phrases that could have been written by anyone, and any tidy three-part list that I did not ask for.

At the end, give me an overall verdict in one honest sentence, and tell me the single change that would do the most to make this sound like me.

Four habits that read as AI

These are worth knowing whether or not you use any of this, because once you can see them you cannot unsee them, and you will start catching them in your own drafts.

Finish it yourself

This is the step people skip and it is the one that matters. A style guide gets you a draft that sounds structurally like you. What makes it actually yours is the material only you have.

1. 🧠 Add your own examples. The specific thing that happened to you, the number from your own work, the thing your manager said. No model can generate these, and they are what makes writing worth reading.
2. ✂️ Cut the paragraph that says nothing. There is almost always one that restates the previous paragraph in different words. Deleting it makes the whole piece sound more human immediately.
3. 🗣️ Read it out loud, then fix what you stumble on. Every stumble is a sentence that is not yours. This takes four minutes and does more than any prompt.
4. 💌 Put one opinion in it. Models hedge by default. A single clear position, stated plainly, is the fastest way to sound like a person with a point of view.

Where the claims come from

• Claude projects (what they are, what each plan gets): Anthropic's projects documentation
• Claude skills (if you want your voice as a portable, reusable file): what are skills
• Detectors and non-native English writers: Stanford HAI, and the underlying paper
• A university switching the detector off: Vanderbilt's statement, and The Markup's reporting on wrongly accused students
• Turnitin's own numbers: false positive rates and AI paraphrasing detection
• OpenAI retiring its own detector: TechCrunch
• Perplexity and burstiness explained: GPTZero