You read your own AI-written email out loud yesterday, didn't you. And about three sentences in you cringed because it didn't sound like you, it sounded like a LinkedIn carousel had a baby with a corporate apology letter. Don't worry. You're in the right place.
The reason almost every AI output sounds like that is not because the AI is broken. It's because nobody has bothered to capture how you write. Not adjectives. Not vibes. The patterns, sentence shapes, word choices, and weird little habits that make your writing recognizable. This process has a name and I call it brand voice fingerprinting. It's the single highest-impact thing you can do with AI right now, and most people skip it.
Here's how I do it. Plus how I package it into a reusable Claude Skill that I can drop into any project.
What "voice fingerprinting" means
A fingerprint is a unique pattern. Yours is different from mine. Same with how you write. You probably:
- start more sentences with "And" or "So" than most people
- avoid certain words on purpose (mine: that corporate L-word that means "use" makes me twitch)
- have a few signature phrases you've been using for years without noticing
- have a sentence-length pattern (mine is short, then long, then medium)
- show up differently in emails than in social posts
- contradict yourself on purpose sometimes
- use parentheses for personality, not only clarification
Voice fingerprinting is the process of cataloguing all of that and turning it into instructions an AI can follow. Then packaging those instructions so you don't have to retype them into every new chat.
Step 1: Pull your raw material
Open a folder somewhere. Could be Google Drive, could be Notion, could be a single Word doc.
Now collect at least 30 samples of your actual writing. I said 30. Not three.
Mix it up. Pull from:
- The last 10 emails you wrote to clients (skip the boilerplate ones)
- Five social posts you wrote yourself (not the ones a VA wrote)
- Three sales pages or long-form pieces
- Five voice memos transcribed into text (this one's underrated, the way you talk leaks into how you should write)
- Five short replies you sent in DMs or Slack
- Anything else that feels like "yeah, that's me"
The reason you need this much material is that AI is a pattern-matching machine. Patterns need data. Three samples isn't a pattern, it's a coincidence.
Step 2: Run the extraction prompt
This is where the magic happens, and it's a long prompt, so plan for it.
Open Claude. (You can do this in ChatGPT too. The output is roughly the same. I prefer Claude for this particular task because in my testing it's been better at picking up on stylistic quirks instead of summarizing tone, but your mileage may vary.)
Paste all 30 samples in. Then ask:
Analyze these writing samples and produce a voice fingerprint document. Include: signature words and phrases this person uses repeatedly. Words they appear to avoid. Average sentence length and rhythm patterns. How they open and close pieces. The kinds of openings they use (questions, declarative statements, observations). Their use of contractions, fragments, parentheticals. Their humor style if any. Their go-to analogies. Their relationship with their reader (formal, friend, mentor, peer). Any forbidden patterns (em-dashes, triplet sentences, etc.) the writing seems to avoid.
You'll get back a document that's probably 1,000-2,000 words.
Read it. Mark the parts that feel right. Cross out the parts that don't. Add anything important the AI missed.
This is now your draft voice fingerprint.
Step 3: Turn it into a Skill
Here's where Claude Skills come in. Anthropic released Skills in 2025 as a way to package reusable capabilities for Claude. A Skill is, in plain terms, a folder with a SKILL.md file plus any supporting docs. You drop it into Claude and it knows when to use it based on the description you give.
For brand voice, I structure the Skill folder like this:
- SKILL.md (the main instructions, including the description that tells Claude when to use it)
- voice-fingerprint.md (the document you produced in step 2)
- writing-samples.md (a handful of your best raw samples, included so Claude can pattern-match against them)
- do-not-use.md (a list of forbidden words and patterns)
The SKILL.md description is the most important part. It's the thing Claude reads to decide whether to load the skill at all. Mine says something like "Use this skill anytime you're writing in Krista Mashore's voice, including blog posts, emails, social posts, ads, and anything else customer-facing." Boring on purpose. Specific.
If you're not technical, a Claude Project with custom instructions can do roughly a lot of what a Skill does, and it's a lot easier to set up. The trade-off is that Skills are portable across projects and chats. Projects are siloed.
Step 4: Test it on three specific outputs
Don't trust your fingerprint until you've stress-tested it. Run it on three things you write often:
- A nurture email to your list
- A LinkedIn or Instagram post
- A reply to an inbound DM or sales question
For each, generate three drafts using your voice skill or project. Read each out loud. (Yes, out loud. Your ear catches things your eye misses.) If it sounds like you, great. If it sounds like an AI doing an impression of you, the fingerprint needs work.
I have a coaching client, we'll call him James. He's a fitness coach. His first fingerprint pass said he uses "lots of metaphors and a friendly, encouraging tone." That description could fit half of Instagram. Useless. The second pass said he opens most emails with a one-sentence story from his week, uses food metaphors specifically (not generic ones), and ends with a question he expects the reader to answer in their head. That's a fingerprint.
The difference between draft one and draft two is the difference between AI that's almost-you and AI that sounds like you walked into the room.
Step 5: Update it forever
A voice fingerprint isn't a one-time project. Every few weeks, when you catch the AI saying something that's not quite right, you go update the SKILL.md or the Project instructions.
I update mine maybe twice a month. Tiny tweaks. A new "don't use this word" gets added. A new pattern I've started using gets documented. (Right now I'm trying to use the word "okay" less and the phrase "so." more. We'll see if it sticks.)
The MCP angle, since you asked. If you have your Claude set up with the Model Context Protocol connected to your Google Drive or Notion, you can keep your voice fingerprint living in one of those places and Claude can pull from it dynamically. Not required. Helpful for teams who need to share the same voice across multiple people.
What changes when this is right
Your editing time drops. A lot. I went from spending 30 minutes editing a 700-word email to about 6. That's a 24-minute swing per piece. Multiply that by the number of pieces you write a week and that's a real, recoverable block of time.
Your content also gets more consistent across channels, which is what builds the kind of trust that turns followers into clients. The kind of authority that compounds.
If you want the full walkthrough on how I built my own voice Skill, including the actual SKILL.md template and the extraction prompts I use, that's all inside my Level Up training. It's the same voice system I use to produce more content per week than most coaching businesses produce in a month. Grab it at kristamashore.com/LevelUp.