For about two years, the way I told a machine to write like me was to describe myself. Punchy. Direct. Short paragraphs. Concrete over abstract. No em-dashes.
That description is a caricature, and a caricature is exactly what came back. Every time. The prose had the silhouette of my writing and none of the weight, the way an impressionist gets the walk right and the person wrong.
The problem was never the adjectives. It was that I was asking the model to imitate a description of me instead of repeating me. Those are different jobs, and only one of them has a ground truth.
So this week I rebuilt the whole thing around a single sentence: the model should repeat the author, not imitate a described voice.
Quoting beats describing
The first half of that is mechanical. Stop telling the model what I sound like and start handing it paragraphs I actually wrote.
My manuscript corpus is 115 documents. Chunked and embedded locally, it becomes 1,353 passages that can be searched by meaning, so when I ask for a scene the prompt arrives carrying six passages that are topically near it and two more chosen purely as voice exemplars. Not a description of my style. My style, in my words, sitting in the context window.
The old behavior was to stuff every chapter in until a 400,000 character budget ran out and the later chapters silently fell off the end. That is not context. That is a truck with the tailgate down.
The retrieval half took a day and is not the interesting part. It is plumbing, and plumbing is below the waterline where it belongs.
The interesting part is the question it forces: how do I know whether it worked?
A voice is a distribution, not an adjective
You cannot answer that with adjectives, because adjectives cannot fail. "Punchy" has no unit. Nothing can be measured against it, which means nothing can ever be shown to have drifted from it.
So I measured instead. Across 27 draft chapters and 116,696 words, my fiction voice came back as a set of numbers. Mean sentence length 6.7 words. Named emotions 0.13 per thousand words. Em-dashes 0.34 per thousand. Dialogue 15.6 percent of the text.
Seeing that was a small shock. Not because any single number surprised me, but because the profile is genuinely distinctive. Two of those figures are near zero, and the near-zero ones are the identity. I do not name what a character feels. I almost never reach for the em-dash. A hundred thousand words of consistently not doing something is a stronger signature than anything I do on purpose.
Then I pointed the thing at a chapter I wrote myself.
The eight
Four hundred and five sentences. Eight flags. A two percent drift rate against my own baseline.
Two percent is the number I was hoping for, because a detector that flags nothing is decoration and a detector that flags everything is noise. Eight is a number a human can actually read.
For contrast I planted a paragraph of unmistakably machine prose in the middle of the chapter. Both sentences lit up, with reasons attached: filter verbs stacked up, a named emotion, a hedge pile, an em-dash. The instrument works. It knows the difference between me on a Tuesday and nobody in particular.
But here is the thing I keep turning over. The detector cannot tell me whether those eight sentences are mistakes.
It knows one fact: those sentences are unlike the 116,688 words around them. Unlike is not the same as wrong. A flagged sentence is either a lapse, where I slid into generic register without noticing, or a departure, where I did something I have not done before and it is the best line on the page.
The numbers are identical in both cases. The judgment is not in the data and never will be.
Don't obey your own fingerprint
Which means the obvious next feature, the one that would demo beautifully, is the one that would wreck the work. Auto-fix all flags. Bring every outlier back to the mean.
Run that on every draft for a year and you have built a machine whose sole function is to make you more like you already were. You would have automated the act of never developing. The fingerprint is a snapshot of who I was across a specific corpus, not a specification of who I should be next.
The right move is much more boring. The flags are amber dotted underlines in the editor with a hover reason. No autofix, no rewrite button, no chat reply explaining my own prose to me. It locates. I decide.
There is a second detector in the same layer that is even blunter. In article drafts it underlines second-person prescriptive framing, the "you should" and "you'll want to" register, and it deliberately stays silent on first person. That one is not measuring my voice at all. It is measuring the absence of one. Advice-column second person is the house dialect of generated prose because it is what you write when there is no specific person doing the writing and no specific thing that happened.
The takeaway
The best thing measurement gives a maker is not a score. It is a shortlist.
Eight sentences out of four hundred and five is a shortlist. It took a fraction of a second to produce and costs nothing to run, and every one of those eight needs a human to look at it and answer a question no amount of corpus can answer: did I mean that?
Some of them I did. That is the only output of this entire system that matters, and it is the one part that stayed with me.