A real, working voiceprint. All 53 answers, each in its own place.
Every model has one house style, so everyone writing with it arrives at the same page. Same brief, two answers below: one model reading nothing, the same model reading one voiceprint. Nothing else changed.
Ask for a draft with nothing attached and you get it: the measured middle-length sentence, the tidy line that closes every paragraph, the courtesy formula standing in front of the answer. It is competent writing, and it is the same competent writing everyone else got this week. What separates your draft from that one is mechanical: how long your sentences run, how far they swing apart, what you refuse to say, how you open. Start a new chat tomorrow, or move to a different tool, and the model starts from its default again rather than from you. That is the gap a voiceprint closes. Nine turns into a long session the file was still landing drafts in the writer’s rhythm, 9 times in 16 against 0 in 8 without it. Whether it fades a little after that we have not pinned down, and a fresh chat when it does is prudence rather than a fix.
This is the draft with a voiceprint. Switch to A for the same brief and the same model with nothing attached.
Hey everyone, I want to share a quick update regarding the newsletter. Starting next month, I'm going to be raising the subscription price from $5 to $8 a month.
When I first launched this project two years ago, my goal was to make it as accessible as possible. Since then, the scope of work and the depth of research I put into every issue have grown significantly. To keep delivering the quality, detailed essays you've come to expect—without taking on outside sponsors or compromising—adjusting the rate is a necessary step.
If you're already an active subscriber, your rate will stay locked in at the original price for the next six months as a thank you for your early support. Thank you so much for being part of this journey and making this work possible. If you have any questions or feedback, my inbox is always open.
Marked in red: a phrase our filter names as an AI tell. None of them is wrong. They are the shapes every model reaches for, which is why this draft reads like four other announcements the same reader saw this week.
Gemini 3.6 Flash, cold, one prompt, no system message. First response, unedited.
I’ve been putting off writing this note for a couple of weeks now, mostly because talking about money always feels a bit uncomfortable. But honestly, here's where I'm at.
Starting next month, I’m raising my rates.
Over the past two years, I’ve put a ton of energy into making sure every person I work with gets deep focus from me without feeling rushed. The dream has always been to deliver depth over volume. To keep doing that—and to give each of you the attention you actually deserve—I need to adjust the numbers.
If we’re already working together, your rate stays right where it is until we wrap up current work. New bookings after the 1st will reflect the new pricing.
I really appreciate you all for trusting the work. If you have questions about how this affects what we're doing, just hit reply.
Same brief, same model, same length — both drafts run to about 150 words, which is what A came to on its own. Gemini 3.6 Flash, one prompt, first response, unedited, and not picked from several attempts. The only thing added was one real voiceprint.
Read on its own, A seems fine. That is the whole problem. Nothing in it is wrong, and nothing in it is anyone. Eight tells in one short draft is the shape the reader has already seen four times this week, in four other announcements. B answers the same brief in the same number of words, gives the news a line of its own, and swings from six words to twenty-eight.
Ask any model to write for you and it reaches for the middle of everything it has read. That middle is an average of millions of people, which is why the writing is recognisable and why it belongs to nobody in particular. A preset moves you to a different point on the same average.
A voiceprint pulls the other way. About fifty-three small choices about how you actually write, kept as coordinates rather than adjectives, written down as one file a model can follow. It describes one writer. It fits the way a signature fits.
A drawing. The measurement is beside it.
One brief, two answers, the same model on both sides. The only thing that changed was whether it was reading a voiceprint. Then we counted.
The card above is one draft. Across 60 generations, three briefs and fifteen runs a side, no voiceprint put zero drafts inside the writer's target rhythm band and the free voiceprint put thirteen there. Permutation test, p < 0.001. Hypotheses were written down before any text existed and two of them failed. The study, including what it could not show →
Both drafts are real and unedited. The default one is a competent answer, which is exactly why it earns a place here.
Everything Timbrel learns becomes one portable, human-readable document, plus a fingerprint that's yours alone.
# Voice Profile fingerprint: 7f3a·9c2e·d011 ## Personality core Warm, plainspoken, a little wry. Leads with a clear idea, then earns it. ## Sentence rhythm Short. Then a longer, breathing line that lets the thought settle. ## Humor Dry, sparing, never at the reader's expense. ## Don't - no corporate hedging - no exclamation points - never "delve" or "tapestry"
Long stretches, short ones, and somebody deciding which
One awkward assignment, run through three real voiceprints. Same AI, same prompt. The only thing that changed was the voice file.
Maren OseiResponse 1 of 3
The newsletter goes to $9 a month on August 1. It’s been $5 since 2021.
The arithmetic is simple. Each issue takes about three days — sourcing the incident, reading the actual NTSB or coroner findings, then redrawing the load path until a non-engineer can see exactly where it failed. At $5 those days were subsidized by my consulting income. In practice I was paying to write this. That’s the failure mode of a hobby that quietly became a second job.
The issues don’t change. Same primary sources, same refusal to run anything I can’t verify against a real report. The back catalog is most of the argument for the new number anyway — read the Hyatt walkway teardown and tell me it was worth less than a coffee.
If $9 is more than it’s worth to you, the archive stays free. No countdown, no farewell discount. The price is just the price now.
Flat assertion, no apology. The em-dash is hers on purpose. She names the false-urgency move and refuses it out loud.
Theo ParkResponse 2 of 3
Three years ago I told you all this thing would take eighteen months. We can laugh about that one now.
So here’s where I’m at. The next build adds the co-op mode basically every email asks for, and doing it right means a small team — an animator and a second programmer — instead of me vanishing into a cave for another year. That’s money I don’t have yet. So the price is going up: $20 to $28 when the co-op patch ships this fall.
If you already own the game, you own all of it — every future update included, the way it was promised on day one. I’m not going to be the guy who quietly edits the promise after people already paid. The new price is for the folks who haven’t found us yet.
Honestly? I went back and forth on even telling you this early. It felt a little like asking permission… But you all funded a stranger’s weird dream about a fishing village that runs on the tides, and you’ve earned the unglamorous money stuff before anybody else does. More soon. Go check on your farm.
Opens on a story and lands the news after the context. Talks straight to “you all.” The trailing ellipsis and the admitted doubt are allowed here.
Nell MercerResponse 3 of 3
The subscription goes to $7 a month. Here’s the part where I’m supposed to apologize for that.
I’m not going to.
Look. You’ve sat through this dance a hundred times — the wounded little essay about tough decisions and the rising cost of doing what we love, performed in the key of a man who’s about to start recycling press releases. Spare me. Spare yourself.
Here’s the part nobody says out loud. I write five of these a week. I read the thing I’m angry about all the way through, in full, before I’m angry about it in public — which, in this business, now counts as eccentric. And I’ve never let a PR person buy a sentence. That’s time I sell to no one. Seven dollars buys it back.
If the column was worth your Tuesday at four, it’s worth it at seven, and you already know which. If it wasn’t, the door’s right there. It doesn’t lock behind you. I’d take a thousand readers who think I’m worth the money over ten thousand who resent the invoice.
A forty-word sentence between three-word jabs. She names the objection, then refuses it on its own line, and punches at the genre rather than the reader.
Most tools stop at tone. Thirteen dimensions go further down: the shape of a sentence, the marks it ends on, how far you stand from a reader, whether you argue or tell a story. The last one is a filter that names the AI tells and strikes them.
PersonalityWho is doing the talking
InvolvementHow close you stand to the reader
Sentence RhythmShort strikes, long breaths
VocabularyThe words you reach for without thinking
PunctuationYour fingerprint in marks
StructureHow a thought gets built
Humor & WitWhere the smile sits, if anywhere
Emotional RangeHow much of you reaches the page
NarrativeStory first, or analysis first
PerspectiveThe seat you write from
ContrarianThe instinct to push back
Cultural CompassThe references you steer by
Anti-AI FilterBanned by name in every fileEvery question is tied to a commissioned research review: stylometry, Biber’s analysis of how writing varies, the psycholinguistics of personality, and how language models actually choose words. Then we tested whether the file that comes out of it does anything. With no file, 0 of 15 drafts landed in the writer’s target rhythm. With one, 13 of 15 did.
Tell us who the voice is for: you, or a brand you're building.
A guided conversation of about 53 quick, human choices. You are picking between real passages, not naming yourself in adjectives.
Your voice file builds in real time as you answer, closer to you with every choice.
Download one clean file and drop it into any AI that takes a system prompt.
The free file is a real one, and it comes out in the formats the major tools actually take. Paying once buys the tooling around it: your own filter, dials that re-cut the file, variants, and a locker if you write in other people’s voices.
Before the number, the honest comparison.
Nobody’s stopping you from building this by hand. What’s genuinely slow is the groundwork underneath it. Why we still think it’s worth skipping →
The number itself is small math: a few hours of your own time and some tokens, against fifty-nine dollars and a file that’s already built.
Either way it comes out plain, editable Markdown, so tinkering stays on the table.
A real, working file, in the box each tool actually gives you.
The same voice, with everything the builder can make from your answers.
Live with the file. Put it in front of the writing you actually do, and see whether the drafts come back sounding like you. If it doesn't do what we've said on this page, reply to your receipt inside fourteen days and we refund it in full, and you keep the file. The whole policy is two paragraphs long.
Writing in other people’s voices? Studio is $199 once and holds ten of them under one key, with an intake link you send to the person whose voice you are writing in — they answer the questions themselves, so you never have to ask a client for an archive they do not have. Each voice also renders as a one-page brief you can hand back to them, and commercial use is stated in the terms. For a whole team sharing one voice, get in touch.
Everything people ask before they build, including the one you're probably thinking.
You can, and if you love to tinker, maybe you should. Writing the prompt is the quick part. The slow part is the research-grounded framework underneath it: which stylometric and psycholinguistic signals actually carry a voice, questions that surface them without asking anyone to describe themselves in adjectives, and an anti-AI filter that names the tells one at a time. That took us several weeks of reading and drafting.
Twenty-nine dollars is what it costs to skip that. Set it against a few hours of your own time plus the tokens, and the comparison is yours to make.
If the appeal is making it fully custom, the file you get is plain, editable Markdown. Take it and tweak it to your heart's content. We did the research scaffolding; you keep total control.
One portable voice file, plain Markdown (.md) and text, plus a unique voiceprint ID. Inside: your style across thirteen dimensions, written as plain instructions a model can follow, your personal anti-AI filter, and your protected signature phrases. You can read every line of it. Here's exactly how it's structured →
No, and that surprises people. You never paste, upload, or type your own prose. Timbrel learns a voice from quick guided choices, mostly "which of these passages sounds more like you?", plus a couple of optional words you love or want banned. Nothing to write, nothing to hand over.
Any that take a system prompt or custom instructions: ChatGPT, Claude, Gemini, Cursor, and most of the rest. It's not a plugin or an integration; it's a file you bring with you, so it isn't locked to one platform. The exact box to paste it into, tool by tool →
Paste it once into the system prompt, custom-instructions, or "project" slot of whatever model you use. From then on it rides along on everything you write there, so you never have to paste it in again.
Like you. It's measured: we map your style from the choices you make instead of asking you to describe yourself in adjectives. It captures your sentence-length swing (burstiness), your involved-or-informational lean, your punctuation habits, and the exact phrases you genuinely overuse, then bans the machine's defaults so they don't paper over yours. See three real voiceprints on the same brief →
Completely. It's plain Markdown with nothing locked. Rewrite a line, add a rule, swap a protected phrase. It's yours to hack. Most people who live in AI tools want to tinker, and this is built to be tinkered with.
There's nothing to collect: you never hand us any writing. We store only your answers (the choices you made) and, if you give it, your email, used purely to generate and send your file. We don't train on anything and we don't sell it. Since there are no accounts, there's no profile of you sitting around to leak.
No account and no password. We do ask for an email at the end, and only for one reason: to send you a copy of the file so it isn't stranded in one browser. You see it on screen either way.
Free gives you the whole thirteen-dimension conversation, your Voice Report, the baseline anti-AI filter, and the file compiled to fit the boxes in ChatGPT, Claude and Gemini. It is a real, working voice file and it is measured: with no file, 0 of 15 drafts landed in the writer’s target rhythm; with the free one, 13 of 15 did.
Full, $59 once, is the same voice with everything the builder can make from your answers attached. A filter built from the phrases you named, on top of the generic list. Your signatures held exempt from it. The passages you picked, copied in as worked examples. A hard-rules block and a numbered check the model runs on its own draft. Dials that re-cut the file. The linter. Unlimited context variants for whatever room you are writing in. Coding-agent rules, a bare system prompt, JSON.
What it is not is a better draft. We measured that too, and on a single draft the free file and the paid one were indistinguishable. You are paying for reach and control, not for a better sentence.
A one-line "sound like me" gives the model nothing to measure, so it falls back on the same averaged defaults. The instruction is still in front of it every turn; it is simply too vague to outrank them. A voiceprint hands it actual coordinates instead. Why "write like me" never quite works →
