My last post told the story: a CEO read my AI-assisted LinkedIn post and told me, kindly, that I'm more interesting than it was. This post is the fix, step by step, so you can do the same thing in an afternoon.
The short version is that you show your voice rather than describing it, because descriptions like "direct and warm" match half of LinkedIn while your own sentences match only you.
Step 1: Gather your real writing
Collect 4 to 6 pieces you wrote yourself, before AI or without it. Mix the registers: a proposal or report, a few everyday emails (one warm, one operational, one where you changed your mind mid-thread), and anything longer you were pleased with.
Two rules here: older is fine, and unpolished is better. My best sample was something I had written years earlier and never tidied. The typos told Claude I write fast and don't buff my sentences, which is exactly the texture AI drafts erase. Don't tidy anything before you hand it over.
Step 2: Create the correction sample
This is the step most people skip and it's the most valuable one. Take something AI drafted for you, and rewrite it in your own hand. Roughly is fine. Then give Claude both versions and ask it to list every change you made and why.
Every edit you made is an instruction. Where you cut a fancy line, that's a rule. Where you replaced a dramatic claim with a plainer, truer one, that's a rule. My rewrite produced the single most useful rule in my whole voice file:
Never let a sentence sound truer than it is.
Step 3: Have Claude write the voice file
Give Claude the corpus and the correction sample and ask for a voice reference document. Tell it to be specific: sentence patterns you use, words you reach for, how your warmth changes between a new contact and an old one, and paired examples of "sounds like me" and "doesn't sound like me" pulled from your real text.
Push back on anything vague, because a vague rule produces a vague draft. "Professional but approachable" is useless. "States the judgement in the first sentence with the reasons attached" is a rule Claude can follow.
One more thing to put in the file: a craft budget. My natural writing is quite dry, and I want AI to make it more interesting, not to impersonate my limitations. So my rule is one crafted line per section, earned, with plain sentences around it. Decide your own number, but set one. The AI giveaway is never one good line. It's ten of them in a row.
Step 4: Package it as a skill
A Claude skill is just a folder. Inside it goes a file called SKILL.md with a short name and description at the top (this tells Claude when to use it), and the voice document either in the same file or in a references subfolder alongside it. Ask Claude to structure this for you; it knows the format.
Zip the folder, then in Claude go to Customize, then Skills, and upload it. You'll need a paid plan with code execution switched on in settings. The exact clicks change from time to time, so check support.claude.com if the menus have moved.
From then on, your voice loads automatically in any new conversation where you ask for something written as you. No re-explaining, ever.
Step 5: Test it cold
Start a fresh chat and ask for something small: a two-line reply, a short email. Read it aloud. If you stumble on a sentence, that sentence is the machine, not you. Feed the correction back and update the file.
Step 6: Have someone who knows you read it
Reading it aloud catches the sentences you stumble on. It does not catch the ones you have stopped hearing, and after a few rounds of editing that is most of them. For those you need someone who knows how you write and owes you no politeness.
For me that was my daughter, who read a post that had already been through the skill and told me two things. The short sentences were not how I write. And where it said "how AI actually works", I would have written "how AI works".
Both were right, and neither had occurred to me, because I had been over that post so many times I could no longer hear it. They are two rules in the file now: one setting a minimum sentence length, one on filler words with "actually" named. The file is not a thing you finish.
Change what you ask for
The six steps above make the drafts better. The bigger lever sits further upstream, in what you ask AI to produce in the first place.
If you want something that is yours, stop asking for finished prose. Ask instead for a research and structure brief: the argument, the evidence, the examples, and a suggested order to put them in. Do the research together, push the idea about until it is right, then write the prose yourself from the brief. Less AI in the input means less of it to strip out at the other end.
Claude's own suggestion for what comes after that, which I am going to try:
After you have done two or three posts this way, give me your spoken versions alongside the original drafts. The differences between them are your voice. I can extract that into a writing-style profile so future drafts start much closer to how you would say it, and the rewrite step shrinks from an hour to fifteen minutes.
That is the correction sample from step 2 again, run on the way I talk rather than the way I edit. I am three posts away from knowing whether it works. It is the next thing I am doing.
The rule that makes it all work
The skill gets Claude 90% of the way there. The last 10% is still yours, and for anything personal it has to be. My rule, written into the skill itself: for anything published under my name alone, I do the final pass by hand. The skill makes that pass a twenty-minute job instead of a rewrite. It doesn't remove it.
Because here's what I learned building this. Twice I cut a line on instinct, got argued back into it by AI, and then had an independent review flag that exact line as the most machine-made thing in the piece. The voice file makes the drafts better. The judgement about what's really you doesn't transfer. That part of the job was never the machine's to do.
An afternoon of setup, and every draft after that starts from your voice instead of a stranger's.
Written with Claude, then steered, cut and checked by me. I answer for it.
Build practical AI skills on your team’s real work. You can book training directly, without taking an index first.