How to Write a Claude System Prompt for Your Voice (With Real Examples)
Most advice on this topic is too vague to be useful. “Tell Claude your tone.” “Give it some examples.” “Describe how you write.” That’s fine as a starting point, but it leaves out the specifics that actually change the output — and without those, you’re still spending half your editing pass correcting back toward your own voice.
I built my own voice profile before I built the tool that generates them for other people, which meant I had to figure out what actually belonged in it through trial and error. What follows is what I learned, with real examples from my actual profile rather than hypothetical ones.
What a Claude system prompt for voice actually is
A voice profile lives in your Claude Project instructions — a settings area Claude reads automatically at the start of every conversation in that project. It’s not a prompt you write each time. You write it once, and from that point on Claude starts every conversation already knowing how you write.
The difference between a useful voice profile and a vague one comes down to specificity. “Be conversational” is not specific. “Don’t use one-sentence-per-line dramatic formatting” is specific. Claude can follow the second instruction. The first one it has to interpret, and its interpretation will be based on what “conversational” tends to look like generically — which is not the same thing as how you actually write.
The five things that belong in it
1. Hard rules — stated in the negative
The most useful section of any voice profile is the list of things you’d never write. These are easier to articulate than positive descriptions of your voice, and they do more work because they’re the things Claude defaults to without them.
Here’s what this looks like in my actual profile:
Never:
- Fabricate anecdotes or invent examples
- Use performed vulnerability (“I used to think X. Now I realise Y.”)
- Write one-sentence-per-line dramatic formatting
- Use “And honestly?” or “Here’s the uncomfortable truth:” framing
- Reference anyone “quietly” doing something
- Use generic inspirational closings (“Your next great story is waiting!”)
- Describe things as “game-changer,” “transformative,” or “life-changing”
- Sound like a guru, motivational speaker, or thought leader
Every item on that list is something Claude would have done without the instruction. Some of them it does constantly — the performed vulnerability structure and the dramatic one-liner formatting show up in AI output so often they’ve become their own identifiable pattern. Naming them explicitly is the only way to get Claude to stop.
2. Anecdote policy
This is the one most people miss, and it’s the one that causes the most jarring failures. If your writing relies on personal stories — and most good writing does — Claude needs explicit instructions for what to do when it doesn’t have one.
Without this instruction, it will invent something. The invented story will be plausible and specific enough to feel real, which makes it worse than a vague placeholder, because you might not catch it immediately.
The instruction I use: Flag where a personal anecdote would strengthen the piece with [add a personal story here about X] rather than fabricating one.
That’s it. One sentence. It prevents an entire category of failure.
3. Confidence calibration by topic
How certain you sound should depend on how much experience you actually have on a topic. Claude doesn’t know this without being told, so it applies a uniform confidence level to everything — which usually means too confident on topics where you’re still learning, and sometimes not confident enough on topics where you have real authority.
In my profile, this looks like: Bold on UX topics — 15 years of experience, can state things flatly. Exploratory on AI topics — still learning, frame as observations rather than conclusions. Genuine on personal topics — write from lived experience, not authority.
The effect on output is significant. With this instruction in place, AI-related content comes out as “here’s what I’ve noticed” rather than “here’s what you should do” — which is both more accurate and more readable.
4. Voice calibration examples
Written instructions only go so far. The most efficient way to teach Claude your voice is to show it the difference between what you’d write and what it defaults to. My profile includes a before/after calibration section that looks like this:
Too cold / corporate:
“Organizations should leverage FAQ optimization to enhance customer experience metrics.”
Too performative:
“Here’s the truth nobody wants to admit: your FAQ page is silently destroying your business. And honestly? Most companies don’t even know it.”
Just right:
“A well-designed FAQ page doesn’t only reduce customer support emails and calls; it is a critical step in the customer journey that can make or break your sales experience. Instead of thinking of help pages on your website as just content, think of them as the first stop for customer service.”
Three examples covering the same territory — one too cold, one too hot, one calibrated correctly. Claude reads all three and understands the target range without you having to describe it abstractly.
5. Real writing samples
This is the section that makes everything else work better. Three to five pieces you’ve actually written, that actually sound like you, teach Claude your patterns faster than any amount of written instruction. Sentence rhythm, paragraph length, how you open and close, when you use headers — all of that comes through in examples in a way that’s hard to capture in description.
Paste them in full if the pieces are short. For longer pieces, pull the opening paragraph and a representative middle section. The goal is enough real text that Claude can detect the patterns, not a complete archive.
How to test it
Once your profile is in place, start a new conversation inside the project and ask Claude to write something short — a LinkedIn post, an email, a paragraph on a topic you know well. Then read it the way you’d edit it: are there phrases you’d cut immediately? Does the confidence level feel right? Does it open the way you’d open something?
The first draft after installing a profile is usually noticeably different from what Claude would have produced without it. If it’s not, the profile needs more specificity — particularly in the hard rules and voice calibration sections.
Building it vs. using a tool
Writing a voice profile from scratch takes a couple of hours if you’re doing it properly — working out your hard rules, finding the right writing samples, writing the calibration examples. It’s worth doing if you want full control over what’s in it.
The Aligned Voice Profile is a fifteen-minute interview that generates the profile for you. It asks the questions you’d need to answer anyway and structures them into a ready-to-install skill file. The output covers all five sections above. You install it in Claude Projects and test it the same way.
Either way, the profile is a living document. As your writing evolves, or as you start using Claude for new types of content, you’ll add to it. The anecdote rule I mentioned was added after I needed it — not before. That’s normal. Start with what you know and refine from there.
Related reading:
What “Write in My Voice” Actually Means to Claude
Why AI Content Never Quite Sounds Like You (And What to Do About It)
How to Get the Most Out of Your Aligned Voice Profile
8 Comments