Telling AI to sound like you

doesn't work. Measuring does.

Browsing LinkedIn used to be fun. Reddit too. Real people writing real thoughts, and you could tell. Then somewhere in the last two years it all started sounding exactly the same. Same confident opener. Same tidy takeaway. The guy who used to write one genuinely thought-provoking quote a week started posting full-blown AI novellas. Boring ones. I don't know if anyone reads the whole thing.

Em dashes everywhere. Emails, product descriptions, half my feed, machine-written, because it was. My dopamine hit from doomscrolling was hitting speedbumps.

I'd love to be smug about it. I can't. I draft with AI like everyone else, and left on defaults, the novella guy is me.

Telling it doesn't work

I tried the obvious fix first. Custom instructions. Remove the em dashes. Remove the 1-2-3 patterns. Remove this, remove that. The model obeyed, and it stripped out important information right along with the tells. Negative rules don't carve the AI out of the writing. They carve the writing.

Then the built-in personalization tools. They change the agent's personality. Friendlier, blunter, whatever you pick. The writing underneath stays too AI and on the nose. Attitude isn't voice.

"Make it sound like me" is the prompt everyone tries. It does almost nothing, because the model has no idea what "me" is. You're aiming it at a target you never gave it.

Majority vote

The strength of a language model is exactly the problem here. It runs on majority vote. That's the training: read everything, then predict the most expected next word. Andrej Karpathy calls a trained model a lossy compression of the internet, and lossy is the word that matters. Compression keeps what's common and buries what's rare.

For most work that's the feature. For sounding like you it's the bug. The vote takes away your grammatical mistakes. Your sentence lengths. The lol, the trailing dots, the abbreviations, the Star Wars reference only three people will get. The thing that makes you is gone, and it's replaced by majority vote.

If you'd rather have it in signal terms: it's a low-pass filter on style. Keeps the frequent, clips the spikes. Your voice is the spikes.

And it's not just my feed. Researchers sat people down to draft with a model and watched different authors start sounding alike. Groups brainstorming with ChatGPT converged on the same few ideas. One team counted vocabulary across fifteen million medical abstracts and found "meticulous" running at thirty-four times its old rate. By 2026 the journals gave the whole thing a name, the homogenizing effect. The studies are all at the bottom if you want them. Funny part: the labs tuned out "delve" once it became a meme, and the fingerprint just moved, to the rule of three, to the em dash. There's always a signature. It changes clothes.

Going to the source

So I stopped iterating on prompts. When something's broken I don't keep poking at it, I go to the source. I have a research pattern I run on everything: an agent scrapes wide across sources, dumps it all into a vault, and I connect the dots. Also, yes, I used AI to find the fix for AI writing.

It came back with something sixty years old. Stylometry. In the early 1960s, Mosteller and Wallace settled who wrote the disputed Federalist Papers by counting tiny words. "Upon." "Of." Not the ideas. The glue. The words you don't notice you're using, which is exactly why you can't fake them. Burrows turned the method into a formula in 2002. Forensic linguists still use it.

Your voice is measurable. And what the math measures best is precisely what the model averages away first: the deviations.

An opinion isn't a measurement

The first version of the tool did the obvious thing. Send the writing samples to a model, ask it to describe the voice. It gave me a different answer every run. A different answer from every model. That's an opinion, not a measurement.

I've spent a decade on systems that punish imprecision. High-precision firmware, where timing is deterministic or the chip doesn't ship. Game loops, where the same inputs have to replay the same frame every time. Same rule everywhere: if you can't measure it the same way twice, you can't trust it.

So the measuring is code. It counts. Sentence lengths and how hard they swing. Fifty-plus function words against English baselines. How wide your vocabulary runs, down to the words you use exactly once. Punctuation habits, trailing dots included. Same input, same numbers, every run. The same counting Mosteller and Wallace did by hand, just faster. A model only shows up at the end, to organize the numbers into a profile you can read, and a second model checks that write-up against the measurements. AI does what it does best, and nothing else.

The easiest number to see is what I call burstiness, how much your sentence lengths swing. I measured my own LinkedIn posts: 0.66. Run the same writing through a model and it drops under 0.3. Flat. A metronome. The flat part is everyone. The swing is the part that sounds like me.

My brother said the output was crap

I tried it on my own emails first. Worked better than I expected. My profile reads like a spec of me: direct, no fluff, data to back it up, short paragraphs you can glance. Now I tell the model to write the damn email and it comes out structured the way I'd structure it. I read it once and hit send.

My emails have names and numbers in them, so here's the demo from the site doing the same move.

Then I gave it to my brother, and he broke it. He's the creative one. He runs a D&D campaign, and his review of the early output was exact: it's crap, he can't use it. He was right. An email profile optimizes for clarity. Steady paragraphs, low heat, function words near business baselines. A campaign voice lives in the spikes, and the profile wasn't catching them.

Email voice and creative voice turn out to be different measurement problems. So the creative path got its own machinery. The interview asks what readers should feel, not what they should do. Characters get measured as separate voices. And the analyzer learns the signature phrases you repeat without noticing, which is how a "Behold" survives a rewrite. Creative voice isn't smoother writing. It's wilder numbers, held on purpose.

Here's where it landed. Same prompt, same model, with and without him in it.

The first one is any AI. The second one is my brother at the table. No model writes "little silver idiot" by default.

One honest limit. The measuring is exact, it's just counting. Getting a model to hold the shape is best effort, and over a long answer it still drifts back toward the middle. But you start from your shape instead of the average, and that turns out to be most of the fight.

How to get it

I named it myDamnVoice because that's exactly what I wanted back. It's free to use, and the promise is simple: measure once, reuse everywhere.

Five minutes on mydamnvoice.com. It interviews you about how you write and what you can't stand, then you paste a few samples and it counts. If your answers contradict your samples, the samples win. The counting runs in your browser, the numbers are deterministic, and your samples are gone the moment they're read. Nothing stored, nothing trained on. If you'd rather stay in your own tool, there's an open-source skill that estimates the same numbers with the model you already use, so your writing never leaves your setup:

npx skills add rangrot/mydamnvoice

Either way you do it once. Paste the profile into your ChatGPT custom instructions, a Claude style, your CLAUDE.md. From then on the model starts from your fingerprint instead of its average. A voice profile is always evolving, but it's deterministic. Re-measure whenever you've changed.

The full science lives on the research page, with a graph of the flattening that says it better than I just did. And the synthesizer check scores any text for flatness, instant, in your browser. I ran this essay through before publishing: burstiness 0.66, zero em dashes, 100 out of 100. It flagged one word, the AI favorite I quoted a few sections up when I called it a tell. Yes, I'm grading my own homework with my own ruler. The ruler is open source. Run this essay through it yourself.

I want my feed back

The averaging won't stop on its own. Next year's models train on this year's averaged output and come out blander still. Competent prose is free and infinite now, so it's worth nothing. The rare thing is writing that sounds like one specific person who actually thought about something.

Intelligence got cheap. Your voice didn't.

Also, my motive here is selfish. I miss reading you. Build the profile, write like yourself again, and give my doomscrolling dopamine a fighting chance.