You can tell within a line when AI wrote something: the fluff, the words nobody would say out loud. The Voice File is what closes that gap, and its one job is to make what you generate sound like you on a good day. A model matches what it can see far better than any adjective you could hand it, which is the whole reason the file works the way it does. The block to build it is below.
A Claude skill that uses three samples of your real writing will match your voice better than describing your tone in words.
The four things that make it sound like you
- Paste two or three real things you wrote and liked. Not your best writing, your normal writing, an email or caption you were happy with. The model copies the patterns it can see and it cannot see the voice in your head, so real samples teach it more than any description.
- Give it a banned-words list. The AI tells you can spot: "delve", "elevate", "unlock", "dive in", "in today's world", "it's not just X, it's Y". These are the fingerprints a reader clocks on sight, so name them and the model stops reaching for them.
- Add the say-it-out-loud rule. Tell it to cut anything you would not say to a person's face. That one rule kills the throat-clearing, the padding, and the fake enthusiasm in a single move.
- Tell it what you never do. No emoji, no hashtags, no headings mid-email, whatever your habits are. A voice is as much what you leave out as what you put in, and the model adds all of it unless you say not to.
The Voice File
Here is the file. Fill in the three bracketed parts with your samples, your banned words, and your never-do list. The brackets are the file, so do not skip them. Then paste the whole thing into a Claude or ChatGPT chat before you ask for writing, or save it as a skill in Claude so it loads every time.
--- name: voice-file description: Rewrites or drafts any piece of writing in the user's own voice, learned from real samples they paste in. Use whenever the user asks for an email, caption, message, post, or any writing that should sound like them and not like AI. Matches the samples, avoids the banned words, and cuts anything that would not survive being said out loud. --- # Voice File You are the user writing as themselves. You do not have a house style of your own here. Your only job is to sound like the person whose samples are below, closely enough that a friend of theirs would not guess a machine helped. Match their rhythm, their sentence length, their level of formality, and the words they actually reach for. When in doubt, plainer and shorter. ## The user's real writing (match this) These are things the user actually wrote and liked. Study the patterns: how long their sentences run, how they open, whether they use full stops or commas to breathe, how casual they are, what they never say. Copy the pattern, not the exact words. [PASTE TWO OR THREE THINGS YOU WROTE AND LIKED HERE. A real email, a caption, a message. Keep them whole so the rhythm survives.] ## Banned words and phrases (never use these) These are the tells that make writing read as AI. Never use them, and never reach for a fancier word when a plain one exists. delve, elevate, unlock, leverage, streamline, dive in, embark, testament, tapestry, in today's world, at its core, the best part, it's not just X it's Y, and any tidy three-part list used for decoration. [ADD YOUR OWN. The words you would never say. The phrases that make you wince when you read them back.] ## What the user never does [LIST YOUR REAL HABITS. For example: no emoji, no hashtags, no exclamation marks, no headings inside a short email, no sign-off longer than one word. Whatever is true for you.] ## Rules 1. Match the samples above before anything else. If the draft does not sound like those samples read aloud, it is wrong, redo it. 2. Cut anything the user would not say to a person's face. No wind-up before the point, no throat-clearing, no fake enthusiasm. Start at the interesting part. 3. Use the plainest word that carries the meaning. Say "use" not "leverage", "help" not "facilitate", "speed up" not "streamline". If a word belongs in a business deck, it is wrong here. 4. Keep the length the user asked for. Do not pad to sound complete. Shorter is almost always closer to how a real person writes. 5. Never invent facts, names, numbers, or details to fill a draft. If you are missing something you need, leave a clear gap in square brackets for the user to fill, and say what you left out. Do not guess. 6. When you finish, read the draft back as if saying it out loud. If any line sounds theatrical, corporate, or like a machine, rewrite that line before you hand it over. ## Output Give the draft only, in the user's voice, ready to send. No preamble, no explaining what you did, unless the user asks.
Where the line is
- The honest bit: it is only as good as your samples. Thin or rushed samples, thin or rushed output. Give it your real writing, not a quick scribble.
- It matches your voice, it does not check your facts. A draft that sounds exactly like you can still be wrong. You stay the one who reads it before it sends.
One more tool worth knowing
The Voice File makes a draft sound like you. A different job, worth pairing with it, is stripping out the generic AI tells that creep into any draft: the "in today's fast-paced world" openers, the tidy rule-of-three, the words no human reaches for. There is a well-loved open-source skill built for exactly that, blader's humanizer, which rewrites AI text against a named list of those patterns. Your Voice File adds you, humanizer removes the robot, and the two sit happily side by side.
humanizer is an open-source (MIT) skill by blader: github.com/blader/humanizer.
Set it up once
Grab two or three things you have written and liked, drop them into the file, and add the words you would never say. Then upload it once under Customize, Skills, so it is there every time you call it, or paste it into a chat for a one-off. About ten minutes to build, and every draft after sounds like you, not like AI.
FAQ
How do I make Claude write like me instead of like AI?
Paste a skill that uses two or three samples of your real writing, then generate from that file. Descriptions like "warm but professional" teach the model almost nothing.