Stop Asking AI to Guess Your Voice. Measure It Instead.

Type “describe my writing style” into any chatbot, paste in a few paragraphs, and you’ll get an answer in under ten seconds: conversational, engaging, direct with a warm undertone. Save it as voice.md, load it into your next session, and the output comes back sounding like every other founder’s.

That’s not a style. It’s flattery with a filename, and it reads like you because it’s vague enough to be anyone.

The adjective problem

Here’s what actually happens when you ask a model to describe your writing: it doesn’t measure anything. It pattern-matches your sample against thousands of “writing style” descriptions it’s seen before and hands back the median. Adjectives are cheap to generate and impossible to verify, so that’s what you get: a label, not a reading.

The tell is that the adjectives never disagree with you. Nobody gets back “your sentences are erratic and you overuse the word ‘basically’.” You get what a LinkedIn bio would say about you if it were trying to be liked. And it’s useless the moment you act on it: “conversational” doesn’t tell a model what to do differently on the next sentence.

Same shape as the memory problem

I’ve written about this failure mode before, in a different costume. In Your Personal Operating System Shouldn’t Live in ChatGPT’s Memory, the complaint was that you can’t read what ChatGPT actually stored about you, only ask it to summarise its own guess. Adjective-based voice.md is the same trick: a model’s guess at its own guess, dressed up as personalisation.

And in Why AI Setups Rot the point was that setups fail silently. Nothing breaks. The voice.md sits there, the output stays technically on-brief while sounding like nobody in particular, and the drift never traces back because the file looks fine. “Confident.” “Direct.” They’re not wrong. They’re just not doing any work.

A fingerprint, not adjectives

So I built the tool I wanted for my own clients: a voice.md generator that measures instead of guesses. Paste in your own writing, a blog post, a long client email, a few LinkedIn posts, and it runs the text through actual analysis, not a chat completion pretending to be analysis.

What comes back is a fingerprint, six figures on the page: word count, average sentence length, rhythm (how much your sentence lengths swing, read as high, moderate, or even), register (casual, semi-formal, or formal, read off your contraction rate), voice (whether you default to “we,” “I,” or neither), and a Flesch-Kincaid reading grade. None of that is a vibe. Every number comes straight out of the text you pasted. It also reads second-person address, punctuation, and signature vocabulary, folded into the generated file rather than listed as tiles.

Why examples beat rules

The fingerprint isn’t even the part doing the heaviest lifting. Models copy a concrete example far better than they follow an abstract rule. Tell one “write shorter sentences” and it drifts back to its house style within two paragraphs. Show it three of your actual sentences and it has something to imitate instead of something to interpret.

The generator filters your sample to sentences between 7 and 34 words, ranks them by distance from your median sentence length, and deduplicates so you don’t get three examples that all open with “I.” Up to three survive and land in the generated file verbatim, with one instruction: write new material that sits next to these. The variety score rewards pasting in a few different pieces over dumping in one giant essay. Range beats volume.

The tell-scanner, the strength meter, and one honest trick

The other half of sounding like you is knowing what you don’t sound like. The tool ships a curated list of AI-tell words and phrases (delve, leverage, robust, comprehensive, “it’s not just X, it’s Y,” “in today’s fast-paced world”) you can ban in one toggle. Before you touch it, the tool scans your own writing and flags which tells already show up in it: uncheck any that are genuinely yours. The more debatable toggles, forced rule-of-three, three-tile grids, ship unchecked by default. Even the ban list doesn’t assume you’re guilty until it’s read your writing.

There’s also a sample-strength score from 0 to 100, built from word volume, sentence count, paragraph variety, and distinctive repeated vocabulary. Under 45 is Thin, 45 to 69 is Fair, 70 and up is Strong. When you’re not Strong, it names the weakest signal and gives the same tip every time: if you’re short on writing, paste a transcript of yourself talking instead. A voice note, a Loom, a podcast clip. Unedited speech is often the truest sample of how you sound.

The analysis runs client-side: your writing sample never leaves the browser tab. (A separate, optional field lower down collects an email for product updates and does post to a server, but it isn’t the writing sample.) Once you’ve got a file, the page gives wiring instructions for four places it needs to live: a Claude Project’s knowledge, a CLAUDE.md at your repo root, ChatGPT’s custom instructions, or <writing_voice> tags for a raw system prompt. Minimum input is 120 words. Aim for 300 or more for a truer read.

Why Most Founders Are Getting 5% From Their AI Tools recommended a hand-written persona.md: your tone, plus the phrases you’d never use. Still a fine five-minute start. This tool is the measured upgrade of that move. Voice is one third of what makes an AI setup work, alongside Context and Workflows, and it’s the third nobody installs properly: “sounds like me” feels soft next to wiring up a knowledge base. It isn’t. It’s the difference between output you ship and output you rewrite.

The tool is free, with no email gate on the analysis, because it’s hour one of how I start every AI Brain install: get the voice measured and wired in before we touch context or workflows. If you want the rest of the install, that’s what the AI Brain engagement builds. If you’d rather talk it through first, book a call. Either way, stop asking a model what you sound like. It doesn’t know. Your writing does.