Why "write in my style" does not work
I have watched a lot of people try the obvious thing. They paste a draft into Claude, type "rewrite this in my style," and get back something that sounds like a competent stranger. Then they add adjectives. "More casual." "More confident." "Less corporate." Each pass moves the text a little, and none of them land on the way they actually write.
The problem is that "my style" is a label, and a label carries almost no information. Casual to me might mean short sentences and no exclamation marks. Casual to you might mean jokes and lowercase. Claude has to guess, and it guesses toward the average of everything it has read that was described as casual. The average is exactly what you do not sound like.
Style, in practice, is a set of observable choices. How long your sentences run. Whether you open with a claim or with context. Which words you would never type. How you push back when you disagree. How you end an email. Those are patterns, and patterns can be written down, handed over, and followed.
That is the whole method. Instead of describing your voice, you document it. I call the document a voice brief, and building one takes about 20 minutes if you have a few pieces of your own writing to work from.
What goes in a voice brief
A useful voice brief is one page, not ten. It covers six things, and each one should be specific enough that a stranger could check a draft against it.
Sentence length pattern
Not "medium." Something like: most sentences are 10 to 18 words, I drop a 3 to 6 word sentence after every long stretch, and I almost never write one over 30 words. The point is that the pattern is measurable.
Favorite openings
Some people open with the conclusion. Some open with a small concrete scene. Some open with a question they are about to answer. Pull two or three real first lines from your samples and include them verbatim. Claude learns more from three real openings than from a paragraph describing them.
Words and moves you never use
This section does more work than any other. Every writer has a banned list, even if they have never written it down. Mine includes "leverage" as a verb, "excited to announce," "I hope this finds you well," and starting a sentence with "Additionally." Also list structural moves you avoid: bullet lists in emails, rhetorical questions, closing with a summary of what you just said.
How you handle disagreement
This is where AI drafts sound most foreign, because the default is to soften everything. If you disagree directly and then give the reason, write that down. If you acknowledge the other view in one sentence and then state yours, write that down. Include one real example of you pushing back if you have one.
How you close
Endings are a fingerprint. Some people end on the next action. Some end on a single short sentence with no sign-off. Note the pattern and the actual sign-off you use, including punctuation. "Thanks, Sam" and "Thanks!" and "Best," are three different people.
Formatting habits
Headers in emails or not. Bold or not. Digits or spelled-out numbers. Contractions or not. Small things, but a draft with the wrong ones reads as off even when the sentences are right.
Step 1: Gather 3 to 5 samples
Pick writing that is genuinely yours and that you were happy with. Not your most polished piece, your most typical one. Sent emails to someone you know well, posts you wrote without help, internal memos, long messages to a colleague. Three samples is the floor. Five is plenty. Past that, Claude starts averaging again.
Strip anything confidential and swap names for placeholders. Paste them into one document with a divider between each. Somewhere between 800 and 2,500 words total works well.
Step 2: Have Claude extract the voice profile
You could write the brief by hand, but it is faster and more honest to have Claude read your samples and report what it sees. Most of us are poor judges of our own patterns.
I am going to paste 3 to 5 samples of my own writing. Describe how I write, based only on what is in the samples. Do not describe how I should write. Do not compliment the writing.
Produce a voice brief with these sections:
1. Sentence length: typical range in words, how often I use very short sentences, the longest sentence you found (quote it).
2. Openings: how I start paragraphs and pieces. Quote 3 actual first lines.
3. Vocabulary: 10 to 15 words or phrases I use repeatedly. Then list common business or online words that appear zero times.
4. Disagreement: find every place I disagree, decline, or correct someone. Describe the pattern. Quote one example.
5. Closings: how I end. Quote the actual sign-offs and last sentences.
6. Formatting: contractions, digits or spelled-out numbers, lists, bold, headers, exclamation marks, emoji, questions.
7. Three things a careful reader would notice that I probably do not notice about myself.
Rules: every claim must point to evidence in the samples. If a section has no evidence, write "not enough evidence" instead of guessing. Keep the brief under 500 words.
Samples:
[paste sample 1]
---
[paste sample 2]
---
[paste sample 3]
Read the output carefully. Some of it will be wrong, usually because one sample is atypical and it over-weighted it. Cut what is wrong. Add what it missed, especially the banned list, because Claude can only report what it saw and cannot see the words you deliberately avoid. Save the corrected version. That is your voice brief.
Step 3: Draft with the brief plus a source-only rule
Now the part people skip. A voice brief tells Claude how to sound. It says nothing about what is true. Hand over a brief and say "write a post about our new pricing," and Claude will produce something that sounds exactly like you and contains details you never gave it. Confident voice plus invented facts is the worst combination, because the errors are harder to spot when the sentences feel familiar.
So every drafting prompt carries two things: the brief and a source-only rule. Claude may only use facts, numbers, names, and claims that appear in the material you provide. Anything else gets a placeholder, not a guess.
Below is my voice brief, then the source material, then the task. Write the piece in my voice, following the brief exactly.
Source-only rule: every fact, number, name, date, quote, and claim in your draft must come from the source material. If you need something that is not there, write [NEED: what is missing] and move on. Do not add examples, statistics, or anecdotes I did not give you.
Voice brief:
[paste your voice brief]
Source material:
[paste notes, bullet points, data, the email you are replying to, or a rough draft]
Task:
[describe what you want: a reply, a 600-word post, a two-paragraph update, and who it is for]
After the draft, list every [NEED] placeholder in one line so I can fill them in.
The placeholders are the point. A draft with four [NEED] tags is a draft you can finish in two minutes. A draft with four quietly invented facts is a liability you might not catch until someone else does.
The same structure works for email, and I covered that version in how to use Claude for email writing. The brief does not change between formats. Only the source material and the task line do.
Step 4: The "where does this not sound like me" review
Even with a good brief, the first draft drifts, usually in predictable places: an opener that is too smooth, a closing that summarizes, one sentence that runs long, a banned word that slipped in. Rather than hunting for those yourself, make Claude audit the draft against its own brief.
Here is my voice brief and a draft. Compare the draft against the brief, line by line.
List every place the draft does not match. For each one, quote the exact sentence, name which section of the brief it breaks, and give a rewrite that fixes only that problem.
Also flag:
- Any sentence over the longest length in the brief
- Any word from my never-use list
- Any opening or closing that does not match my patterns
- Any sentence that sounds generically polished rather than like me, even if you cannot point to a rule it breaks
Do not rewrite the whole draft. Do not praise what works. Just the mismatches, in order.
Voice brief:
[paste your voice brief]
Draft:
[paste the draft]
This catches the obvious misses, and over time it teaches you what your brief is missing. If the review keeps flagging a problem the brief did not cover, add a line to the brief. After three or four rounds, the brief is tight and the review comes back nearly empty.
Step 5: Store it in a Project so you stop pasting
Pasting a 500-word brief into every chat gets old in about a day. The fix in claude.ai is a Project. Projects give you a workspace with its own instructions and knowledge files, and according to Anthropic's help center, "Claude will use these instructions for all the chats within the project." Projects are available on free accounts too, with a limit of five projects on the free plan.
Create a Project called something like "Writing as [your name]." Paste the voice brief plus the source-only rule into the project instructions, so both apply to every chat automatically. Then upload your samples as a knowledge file, so Claude can refer back to real examples rather than just the summary. Every new chat inside that project starts already knowing how you write.
One caveat from the same help article: context is not shared between chats in a project unless it is in the knowledge base. If you refine the brief mid-conversation, go update the project instructions, or the improvement lives in one chat and dies there.
Through the API or a tool built on it, the same idea applies: brief in the system prompt, samples as reference material, source-only rule next to the brief. If you are still deciding which assistant to standardize on, I compared the two big options in Claude vs ChatGPT for writing, and the method transfers to either.
A prompt for when you have no samples yet
Occasionally someone wants a voice for a format they have never written in. You cannot extract a pattern from nothing, but you can build a starting brief by interview and correct it as real samples accumulate.
I want to build a voice brief for [format: a weekly newsletter, LinkedIn posts, client update emails] but I have no samples yet. Interview me to build one.
Ask me these one at a time, waiting for my answer before the next:
1. Paste one piece by someone else that sounds close to how I want to sound, and what I like about it.
2. Paste one piece that sounds like how I do NOT want to sound, and what bothers me.
3. Do I open with the point or build up to it?
4. When I disagree, do I say so directly or acknowledge the other side first?
5. Contractions, digits, lists, bold, emoji, exclamation marks: yes or no for each.
6. Five words or phrases I never want to see in my writing.
7. How I end: a next action, a question, a single short line, or a sign-off.
Then produce a draft voice brief in the same seven-section format an extraction would produce, and mark every section "provisional" so I replace it once I have three real samples.
Treat that brief as a placeholder. The moment you have three real pieces you are happy with, run the extraction prompt on those and replace it. A brief built from your actual writing will always beat one built from your intentions.
The limits, stated plainly
This method works, and I use it every day, but there are three things it cannot do.
A voice brief cannot make a claim true. It controls how the sentences sound, and nothing else. If your source material contains an error, the draft repeats it in your voice, which makes it more convincing, not less. The source-only rule keeps Claude from adding new errors. It does not fix the ones you gave it. Check your facts before they go in.
A voice brief cannot replace editing. It gets you from a stranger's draft to one that feels mostly you, and the last stretch is still your job. Read every draft out loud. You will hear the places it is not you faster than any review prompt will find them. If you want more control over the raw first draft, the ghost Claude prompt pairs well with this: use it to get a clean skeleton, then the voice brief and review pass to make it yours.
And a voice brief drifts. The one you build in September 2026 will describe someone slightly different by next spring. Re-run the extraction on fresh samples every few months and compare. Usually two or three lines change. Sometimes you learn you have picked up a habit you would rather drop.
Keeping the prompts together
The four prompts above, extraction, drafting with the source-only rule, mismatch review, and the no-samples interview, are the complete kit. I keep mine in a text file next to the voice brief so I can rebuild the Project in minutes if I need to. If you want more patterns in that format, the Claude Cheat Sheet is 120 prompt patterns with worked examples and notes on when each one does nothing, $19 once, two free samples on the page.
If you are newer to Claude and want the fundamentals first, start with the free 75-page Claude guide. It covers prompt structure, context, and catching invented facts, all of which this method leans on. Once that is comfortable, the cheat sheet is the next step if you want a larger library of patterns.