How to Write Prompts for AI Copy That Sounds Like You
Your team drafts a LinkedIn post in ChatGPT. Everyone reads it, everyone nods. Nobody argues it's bad.
Nobody says it sounds like you either.
You edit it. You push some sentences around. You add a specific number. You cut the third bullet. The result is closer, but not really closer. It sounds like a competent generalist who has read your website. Which, technically, is what wrote it.
What most people get wrong about AI copy voice
The most common fix is more adjectives.
"Write in a friendly, professional, confident tone." "Match a direct, no-fluff voice." "Sound authoritative but approachable."
The model reads those, translates them into the average of every "friendly professional confident direct" corpus it's ever seen, and hands back something that reads like a LinkedIn thought-leader who has smoothed off every edge. Which is exactly what you asked for.
The issue isn't the model. The issue is that voice doesn't live in adjectives. It lives in choices. Which words a person reaches for by default. Which sentence lengths they trust. What they refuse to say. What they'll say bluntly that a generalist would soften.
Adjective-based prompting asks the model to average across a corpus. Voice is what happens when a specific person diverges from that average in a specific direction, on purpose, every time.
The mechanism: every prompt moves three axes
Any prompt you write to an AI writing tool is doing three things at once, whether you name them or not:
Task. What the copy needs to accomplish: a subject line, a landing page hero, a cold email opener. Task is the easy part; most prompts get it right.
Constraint. What the copy can't do: max length, must include a link, must not name a competitor. Constraints are the second-easiest part. Most prompts include a few.
Voice. How the copy sounds when a specific person reads it: cadence, word choice, argument structure, what gets emphasized, what gets cut. Voice is where most prompts stop giving the model any signal.
When you don't specify voice, the model has to pick one. It picks the average, because averaging is what it was trained to do when the signal is missing. The average sounds competent. The average never sounds like you.
Voice is a personality signature
Here's the part that took me 20 years of writing B2B copy to see clearly: voice isn't a style, it's a personality signature.
Every writer, including you, makes the same kinds of choices in a consistent direction. Those choices map to the same five personality dimensions psychologists use to describe people generally. The Big Five. OCEAN.
- Openness decides whether a piece of copy reaches for novelty and metaphor, or reaches for structure and precedent.
- Conscientiousness decides how many specific details the writer trusts you to want: numbers, steps, dates, sources.
- Extraversion decides whether the copy warms up quickly with the reader or holds a professional distance.
- Agreeableness decides whether the writer frames themselves as part of a "we," or stands alone as an "I."
- Neuroticism — the safety-versus-gain axis — decides whether the copy leads with what's at risk or what's possible.
Every sentence you write is a small dial-turn on all five. Your voice is the pattern those dials make when you write on autopilot. Your team's voice is a version of that pattern that survived committee review.
You can't get an AI writing tool to reproduce that pattern by describing it in adjectives. But you can get it to reproduce that pattern by naming which dimensions to activate, and which to suppress.
Five ways to invoke voice in a prompt
Here's the shift: instead of prompting for adjectives, prompt for dimension activation.
Openness — novelty vs. structure.
- Weak: "Sound fresh and original."
- Strong: "Open with a metaphor or unexpected reference before you explain the concept." (High-O) OR "Open with the same three-column structure you'd use in a checklist. Skip the metaphor." (Low-O)
Conscientiousness — detail vs. big picture.
- Weak: "Be specific and clear."
- Strong: "Every claim in this post carries a number or a source. If you can't ground a claim, cut it." (High-C) OR "This piece paints the vision. No numbers, no citations, no year-over-year. Just the argument." (Low-C)
Extraversion — warm vs. cool.
- Weak: "Be approachable but professional."
- Strong: "First person, contractions, direct address to the reader in the first paragraph." (High-E) OR "Third person or 'the team.' No 'you' address until the CTA. Reserved register." (Low-E)
Agreeableness — collective vs. individual.
- Weak: "Sound collaborative."
- Strong: "Every declarative is 'we.' No 'I.' No 'my.' The team is the actor throughout." (High-A) OR "Every declarative is 'I.' Personal experience only. Team gets mentioned once at the close, if at all." (Low-A)
Neuroticism — safety vs. gain.
- Weak: "Be motivating."
- Strong: "Lead with the risk of not acting. Every paragraph names what breaks, what leaks, or what stalls if the reader doesn't move." (High-N) OR "Lead with what becomes possible. Every paragraph names an outcome, opening, or upside." (Low-N)
The pattern: replace adjective ("original," "specific," "approachable") with operational instruction ("open with a metaphor," "every claim carries a number," "first person, contractions, direct address").
Operational instructions are things the model can execute deterministically. Adjectives are things the model has to average.
Before/after: the same task, two prompts
Take a real B2B task: draft a subject line for a webinar re-engagement email.
Generic prompt:
Write a subject line for a webinar re-engagement email. Tone should be friendly, professional, and confident.
Output the model tends to produce:
"Don't miss out on our exclusive webinar this Thursday"
"Last chance: unlock actionable insights from industry experts"
"You're invited: transform your workflow in 45 minutes"
Competent. Averaged. Wouldn't fit any of your buyers in particular.
OCEAN-anchored prompt (targeting a High-C, moderate-N reader, a data-driven PMM who scans by structure and weights risk):
Write a subject line for a webinar re-engagement email. The reader is a director of product marketing who registered but didn't attend the live session. They read subject lines by looking for structure and specificity first. They'll ignore anything that reads as marketing pressure.
Constraints: max 50 characters. Lead with a specific number or artifact. No exclamation marks. No "don't miss." No "last chance." No "unlock." The subject line should imply a concrete thing they'll receive, not an experience they'll have.
Output the same model tends to produce:
"The 7-slide summary you asked for"
"Thursday's webinar → the checklist + recording"
"Missed Thursday? The 3-page recap is here."
Same model. Same task. The second prompt named who the reader is in operational terms and forbade the vocabulary that signals low-C generic marketing pressure. The output shifted accordingly.
Two things you'll notice:
- The task is identical. The voice is what changed.
- You didn't teach the model anything new about the reader. You gave it a personality frame it could execute against.
The audit habit
After a few weeks of prompting this way, the useful question stops being "does this AI copy sound like me?" and starts being "can I name which OCEAN dimensions the copy is activating?"
If you can't, the copy is averaged.
If you can, and the dimensions match your reader, the copy is targeted.
If you can, and the dimensions don't match your reader, you have a diagnosis, not a mystery.
The mechanism is scannable. The fix is a re-prompt, not a rewrite.
Where to check your work
If you've written a piece of AI copy and you're not sure which OCEAN dimensions it's activating, paste it into the free Ad Copy Analyzer.
Try it: Free Ad Copy Analyzer
3 free analyses. No login. No credit card.
→ semalytics.com/tools/ad-copy-analyzer/
It scores four dimensions (Engagement, Personality fit, Strategic Clarity, and Framing Strategy) grounded in 860+ peer-reviewed papers in personality psychology. You'll see which dimensions your current copy activates and which it doesn't. That's the diagnosis that tells you what to change in the prompt.
If you want to see the mechanism applied to a different B2B surface, the 5-experiment framework for isolating OCEAN dimensions in subject lines is the sister post.
If you want the full framework, how the four scoring dimensions connect to the underlying research and how COS uses personality signal to generate B2B copy tailored to each buyer, start at semalytics.com/ai-copywriter/.