There’s a specific voice that most AI-generated content defaults to. You’ve read it a hundred times. Here’s why it happens and what two years of building identity prompts across ten different sites taught me about fixing it.
For the past two years, I’ve been doing something most people skip: building writer identity prompts from scratch, testing them across real content, throwing them out, rebuilding them, and refining them until they produce output that’s actually distinguishable from everything else online.
Right now I’m running close to ten different websites — different audiences, different clients, my own personal brands — each with its own custom AI writer profile. Some are B2B consulting voices. Some are agency institutional voices. Some are local, community-facing tones. Each one took iteration to get right.
What I found, consistently, is the same problem showing up everywhere before the fix: a voice that sounded like no one in particular. Competent, correct, and completely generic.
I started calling it universal narrator voice.
What universal narrator voice actually sounds like
You know it when you read it. It opens with “In today’s rapidly evolving landscape…” It uses words like seamless, robust, and empower without any awareness of how hollow they’ve become. It wraps up with “In conclusion” and a tidy three-point summary.
It isn’t wrong. It’s just indistinguishable.
The interesting thing is that I could recognize this immediately — and I think the reason is background. Between 2012 and 2020, I wrote over 10,000 pieces of content manually, across a range of industries that probably sounds implausible: e-commerce, digital marketing, social media strategy, design software tutorials, music and instrument guides, real estate, computer hardware, and more. All of it researched, all of it written from scratch, each sector with its own vocabulary and conventions and reader expectations.
That experience gave me a calibrated sense of what genuine sector voice actually feels like — the specific way a music educator writes versus a hardware reviewer versus a real estate professional. When AI produces content that ignores all of that and defaults to the same averaged-out professional register regardless of topic, you notice it immediately if you’ve spent enough time in the real thing.
The model isn’t making a mistake. It’s making a default.
Why the default exists
When you ask AI for content without telling it who’s writing, it doesn’t guess your voice. It reaches for the most statistically common version of “professional writing” in its training — which is to say, it becomes a universal narrator. Authoritative enough to sound credible, neutral enough not to offend anyone, generic enough to fit any context.
There’s a useful way to think about this: AI knows as much as the language you give it. The more you feed it — the more specific, the more characterful, the more opinionated — the more its output shifts to match that register. The model is, in a real sense, responding to you in the language you speak to it.
Most people speak to it in task language: write a post about X, draft an email about Y. Task language produces task output. The universal narrator fills in the rest.
The instruction most people skip
The fix isn’t a better task prompt. It’s a clearer identity behind the task prompt.
Before you ask for content, tell the model who is supposed to be writing it — not a job title, but an actual characterization. How this person thinks. What they refuse to say and why. What their relationship to the reader actually is. The things they’ve seen that someone without their background wouldn’t have seen.
Think of it as a brief. A detailed brief. The kind you’d give a copywriter you were hiring to write in your voice — not “professional tone,” but the specific qualities that make your professional tone different from anyone else’s.
When I started building these systematically, the change in output was immediate. Same task prompt — “write a LinkedIn post about why most companies wait too long to address their pricing strategy” — produces completely different results when the model knows it’s writing as a 20-year B2B advisor who’s seen that mistake across dozens of clients, versus a first-person practitioner writing from their own recent experience, versus an agency voice publishing a collective institutional perspective.
Same task. Completely different people saying it.
Why this is harder than it sounds
Here’s the catch: most people can describe what they want to say. Far fewer can describe how they sound.
Voice is something we do, not something we typically analyze. Ask a good writer to characterize their tone and they’ll usually give you something vague — “conversational but professional,” “direct,” “friendly.” These aren’t wrong, but they’re not specific enough to give a language model anything to work with. The universal narrator is also “conversational but professional.”
What actually differentiates voice is more granular: the things you don’t do, not just the things you do. The words that would feel false in your mouth. Whether you lead with observation or instruction. Whether you’re willing to make a strong claim or hedge everything. Your relationship to uncertainty.
Getting this on paper — in a form the model can act on consistently — takes a bit of work upfront. But once it’s done, it compounds. The same identity prompt works across dozens of pieces of content. You stop starting from zero every time.
A starting point
If you want to try this immediately, the simplest version:
Before your next request, add two things to the top of your prompt:
1. Who is writing this — not a title, a characterization. “You are a [role] with [X] years of direct experience in [field], writing for [specific audience]. You write from lived observation, not theory.”
2. What this person avoids — the specific things that break the voice. “Never use motivational language. Never open with a rhetorical question. Never use ‘in today’s world’, ‘seamless’, or ‘it’s important to note’.”
That alone will change your output noticeably. Not because you’ve unlocked a trick, but because you’ve given the model an actual person to write as instead of a gap to fill with defaults.
The deeper principle: AI responds in the language you give it. Give it a generic brief, and it produces a generic output. Give it a specific identity — a real characterization of a real voice — and it has something to actually imitate.
That’s the trade-off: a small upfront investment in self-definition for a significant ongoing improvement in everything that comes after.
I’ve spent two years building and refining writer identity prompts across different sites, audiences, and sectors. I put the core of what I’ve learned into a free resource — ten fully written identity prompts for different types of creators and businesses, each one ready to copy-paste into ChatGPT or Claude. You can get it at here.
