Make any AI write in your voice.
Paste in a few things you have written. This builds you a voice.md file:
a short context file you drop into Claude, ChatGPT, or any AI tool so it stops sounding like a
press release and starts sounding like you. The rules are measured from your actual writing,
not guessed. Nothing you paste leaves this tab unless you ask for the kit by email.
Build your voice.md
Say who each sample was written to. How you prompt an AI is a different register from how you write to people, so anything written to a tool is measured separately and kept out of your fingerprint.
Sharpen it with your own edits Optional
Real before and after
Paste something an AI drafted for you and what you actually sent. The tool works out what you changed (the cut closing line, the qualifiers, the length) and turns each habit into a rule. The pair itself goes into the file as an example.
Times you told an AI its draft was wrong
One per line, in your own words: "too salesy", "this assumes I was in the meeting", "it's just a list". They get grouped into taste rules, most frequent first, with your words quoted as the evidence.
Runs in your browser. Nothing you paste is stored or logged, and nothing is sent unless you ask for the kit by email.
Short on writing? A transcript of you talking (a voice note, a Loom, a podcast) works brilliantly here. It captures how you actually sound.
Past 1,500 words of writing to people, the file also lists the common phrases you never use.
Your voice fingerprint
Measured from your samples in this browser. This is what the file below is built from.
Who you write for Optional
Example sentences to keep
Real sentences from your writing. AI copies these far better than any rule. Keep the ones that sound most like you.
AI tells to ban
The cliches that give AI away. Flagged ones already appear in your samples: uncheck anything that is genuinely your habit.
Your own rules Optional
Out of date: your samples changed after this was generated. Analyse again before you copy, download or email it.
Prove it works, blind
Don't take the file on trust. The button below copies a single prompt: your voice.md plus one writing task. Your AI writes it twice, once ignoring the file and once following it, shuffles the two, and asks which sounds more like you. Only after you pick does it reveal which was which. A labelled A/B tells you which one to prefer. This doesn't. It runs in your own AI tool, nothing sent from here.
Or have the kit land in your inbox
Your voice.md plus the blind taste-test prompt, emailed now. It sends the file exactly as shown, so any edit pairs or critiques you added go with it. The email also has a link to join the drift digest, a short note when new AI tells land and a quarterly nudge to re-measure. You're only on it if you confirm through that link, and we never keep a copy of the file.
Drop it into your AI tool
Same file, four homes. Pick yours.
Claude Project / Claude.ai
Add voice.md to Project knowledge. In the instructions: "Before writing, read voice.md and match it exactly."
CLAUDE.md (Claude Code)
Save voice.md at the repo root and add to CLAUDE.md: "For any prose you write, read @voice.md first and follow it."
ChatGPT
Settings → Personalization → Custom instructions. Paste the file into "How would you like ChatGPT to respond?"
Any API / system prompt
Wrap the file in <writing_voice> tags in your system prompt and end with "Match writing_voice in everything you write."
Check a draft against your voice
The generator is a one-off. This is the habit: before you publish anything, paste the draft here and get two readings. Which AI tells it carries, and how far it drifts from your own fingerprint, measured when you generated your voice.md above. The fingerprint lives in this browser only; like everything else on this page, nothing you paste goes anywhere.
Why this beats a paste-and-pray generator
Most voice.md tools send your writing to a server, run it through a language model, and hand back a paragraph of vague adjectives: "conversational, engaging, authoritative". That is a horoscope. It fits everyone, so it captures no one.
This one measures. It counts your sentence lengths and the rhythm between them. It finds the words you actually reach for. It reads your grade level. Then it pulls out real sentences you wrote and tells the AI to write things that would sit next to them, because a concrete example beats an abstract rule every time. The last layer strips the AI tells: the em dashes, the "it's not just X, it's Y", the "delve" and "leverage" and "in today's landscape". You end up with a file that is short, specific, and unmistakably yours.
It is the same first move I make when I install an AI Brain for a founder: get the voice out of your head and into a file the tools can always see. This is that file, generated in about a minute, kept entirely in your browser.
Voices drift. The Drift Digest is the nudge.
Your voice shifts as you write more, and the AI tells change every time a new model ships. The Drift Digest is a short email when there's a real reason to come back: new tells added to the scanner, sharper analysis, and a quarterly reminder to re-measure your fingerprint. No sequences, unsubscribe anytime.
What is a voice.md file?
A voice.md file (you will also see it written as voice md or md voice) is a plain
text file, in Markdown, that describes how you write. The .md is just the Markdown
extension. Claude, ChatGPT, Cursor and other AI tools read it as ordinary text.
It holds your sentence rhythm, your register, the words you actually use and the AI habits you want gone. Given that file before it writes, the AI has something specific to match instead of falling back on the default voice.
It is easy to confuse with two other things:
- voice.md is only how you sound. It works in any tool.
- CLAUDE.md tells Claude Code how to work in a repository. A voice.md can sit beside it and be referenced from it.
- Custom instructions are the settings box in ChatGPT and Claude.ai. Paste your voice.md in there if the tool has no file upload.
Voices drift, so a voice.md is not a one-off. Re-measure it every few months from newer writing.
Common questions
What is a voice.md file?
A voice.md file is a short context file that tells an AI tool how you write: your sentence rhythm, how formal you are, the words and phrases you actually use, and the AI cliches to avoid. You drop it into Claude, ChatGPT, or any AI agent as context, and it writes in your voice instead of the default AI voice. It pairs naturally with a CLAUDE.md or a Claude Project.
How is this different from other voice.md generators?
Three ways. First, the rules are measured from your own writing in your browser, not guessed by a language model, so they are accurate rather than flattering: sentence rhythm, softener rates, and the common phrases you never use. Second, it learns from evidence of your taste, not just your prose: real sentences from your samples, pairs of what an AI drafted against what you actually sent, and the times you told an AI it was wrong. Third, it ships a built-in AI-tell ban list that also scans your text, so you can strip the cliches without banning a habit that is genuinely yours. Nothing you paste leaves the tab unless you ask for the kit by email.
Is my writing stored or sent anywhere?
Not unless you ask for it. The whole tool runs in your browser. Your samples are analysed locally, the voice.md is built locally, and nothing is uploaded, logged, or saved. Close the tab and it is gone. The one exception is the optional "Email me my kit" button: it sends the generated voice.md exactly as shown, which includes the example sentences you kept and any edit pairs or critiques you added, but never your raw samples, and we do not keep a copy of the file.
Why does it ask who each sample was written to?
Because how you talk to an AI tool is a different register from how you write to people. Prompts are full of commands and phrases like "let's" and "make sure"; in one founder's real year of writing, "let's" ran at 64 per 10,000 words in prompts and about 4 in Slack and LinkedIn. Mix the two and the voice file teaches the AI your prompt register. Samples written to a tool are measured separately, and a sample marked as written to people that still reads like instructions gets flagged.
How much writing should I paste in?
Aim for 300 words or more, ideally two or three real pieces you are happy with: a blog post, a long email to a client, a LinkedIn thread. The tool works from 120 words, but more samples give a truer fingerprint and better example sentences. Past 1,500 words it also lists the common phrases you never use, which is only meaningful on a sample that size. Paste finished writing, not notes.
How do I keep AI output sounding like me over time?
Check drafts before you publish them. The drift checker on this page takes any draft and reports two things: which AI tells it carries, and how far it drifts from your own measured fingerprint across five dials (sentence length, rhythm, register, reading grade, and direct address). Your fingerprint is saved in your browser when you generate a voice.md, so the check takes seconds. Voices also shift, so re-generate the voice.md every few months from fresh writing.
What is the difference between voice.md, CLAUDE.md and custom instructions?
A voice.md holds only how you write. A CLAUDE.md tells Claude Code how to work in a repository: build commands, conventions, what to avoid. Custom instructions are the settings box in ChatGPT and Claude.ai. They overlap in practice, so the simplest setup is one voice.md that you reference from a CLAUDE.md or paste into custom instructions. Keeping voice in its own file means you can reuse it in every tool.
Does a voice.md work with tools other than Claude and ChatGPT?
Yes, with any AI tool that accepts instructions or a file as context: Cursor, Gemini, Copilot, or your own app through a system prompt. It is plain text, so nothing in it is tool-specific. Paste it in or upload it, and tell the AI to read it before it writes.
Want the voice wired into everything, not just one chat?
A voice.md is the first file. The AI Brain install puts it, plus your context and workflows, where every tool and teammate can reach it, so nobody pastes anything ever again.
Rather build that yourself? The Founder Course is a 6-week DIY path: same install, your own hands.