Teach Claude to Write Like You, Not Like AI

Why AI writing sounds like AI

Try the obvious road first and describe your tone. You will write something like friendly but professional, direct but warm, simple but smart. Now notice that every freelancer on earth describes their tone in those same words. Claude has read millions of friendly but professional writers, so when you hand it that phrase it returns the average of all of them, and the average of everyone sounds like no one.

Your real voice hides in the mechanics you have never noticed. How long your sentences run before you stop them. Which small joining word you reach for when one thought leans on another. Which plain word you pick when ten fancy ones are sitting right there. Whether you open with a little story or a blunt claim. You cannot type those habits into a prompt because you do not know you have them, and that is the exact reason descriptions fail while real samples work.

Describe a friend's face to a stranger and you get round face, short hair, glasses, which fits half the city. Hand over one photo and the stranger finds your friend in a crowded airport in seconds. A tone description is the description. Your ten old posts are the photo.

A description matches a million writers. Ten real posts match exactly one.

Show your voice, do not describe it

So you stop describing and start showing. The whole training setup is two things sitting in one place. First, ten real pieces of your own writing. Old posts, a long message where you explained something to a client and it landed, a proposal that won. Claude learns from the words themselves, so paste the full text, never a link or a screenshot. This is the one thing you can do today, before anything else, and it does most of the work on its own.

Second, a short voice document, one page of hard rules the samples cannot state by themselves. Both go into a Claude Project on claude.ai, and every chat you open inside that Project inherits them without you pasting your voice again. That permanence is the point. Feed it once and every future draft starts from your rhythm instead of the model's.

The voice document is where most of the power hides, so it earns care. Mine holds a list of banned words, the ones models adore and I would never say out loud. It holds the sentence shapes I refuse and the ones that are truly mine. It names my register, one friend over tea, easy English, no lecture. I am not going to paste the whole file here, because a half copied list teaches Claude the wrong lessons, and the finished version took months of broken drafts to earn.

Read every draft for the three tells

Even with a loaded Project, model habits leak back in, so every draft I keep gets read against three layers. I call them the three tells, and once you learn to see them you cannot stop seeing them in half the posts on your feed.

The first is tell words. A family of words shows up in almost every AI paragraph and lands nowhere near how people actually talk. Here is the free check you can use today, on any sentence, with no tool at all. Read it aloud and ask whether you would say that exact word to a friend over tea. Would you really say realm out loud across a table? If the answer is no, the word is a costume, and readers smell costumes.

The second is tell shapes, and these are sneakier because every single word looks innocent. The pattern is what gives it away, like the not this but that flip landing five times in one page, or every point arriving in a neat group of three because three sounds wise even when the honest count was two. One flip in a whole piece can hit like a punch. Five turn your writing into a stage act.

The third is tell rhythm. Model sentences tend to land at the same length one after another, and the paragraphs stack like identical bricks. Human rhythm wobbles, because a real person rides a long thought that keeps leaning forward until it finally lands. Then stops dead. Short line. That wobble is most of what people mean when they say a piece feels human.

Strip the confident fiction before it ships

The tells are a style problem. This next one is a truth problem, and it is worse. Give a model your samples and it will copy your rhythm faithfully, then decorate the draft with things that never happened. A client name you never had. A percentage you never measured. A small Tuesday that never existed. And it never hesitates. It invents with full confidence, which is what makes the fake line so hard to spot, because it reads exactly like something you would really say.

So after the style read, I run one more pass that checks only truth, kept separate on purpose, because a style pass and a truth pass compete for attention and whichever one runs second gets done badly. The move is plain enough to start today. Ask Claude to list every factual claim in the draft, every number, every name, every event, as a bare checklist. Then walk the list and strike anything you cannot personally point to. Where an invented number falls out, put in the real small one or go fully qualitative, because a true small number survives questions and a made up big one collapses at the first reply.