AI Personas and Brand Voice: How to Stay You While Scaling

Nobody talks about this enough. The biggest risk with AI-generated content isn’t that it’ll be bad. It’s that it’ll be fine. Competent. Forgettable. Like every other AI-generated piece of content your competitors are also churning out.

I’ve seen smart brands go from having a distinct personality to sounding like a LinkedIn post generator because they handed their content over to AI without doing the hard work first. The hard work is defining what your voice actually is, with enough specificity that an AI can approximate it.

Here’s how I do it and what I’ve learned.

Your Voice Is Made of Specific Choices, Not Adjectives

Most brand voice guides I’ve seen are useless for AI prompting. They say things like “we are friendly, professional, and approachable.” Great. So is literally every other brand. That description produces nothing distinct.

Voice is not adjectives. Voice is specific choices.

• Do you use contractions? Always, sometimes, never?
• Do you reference personal experience in the first person?
• Do you swear, and if so which words and in what contexts?
• Do you use analogies? What kind?
• How long are your sentences on average?
• Do you use rhetorical questions?
• What topics do you never comment on?
• Who do you explicitly agree or disagree with?
• What phrases are banned because they sound like everyone else?

Answer those questions and you have something an AI can actually work with. The adjectives are nice for humans to read. The specific choices are what an AI can implement.

Build a Voice Guide, Not a Style Guide

A style guide tells you formatting rules. A voice guide tells the AI who it’s pretending to be. They’re different documents.

My voice guide for my personal brand is about 500 words. It covers:

Who’s speaking: Not a corporate brand. A specific person with a specific background, specific opinions, specific pet peeves.

What the reader relationship is: I write like I’m talking to someone I respect who’s capable of handling direct information. Not like I’m explaining to a beginner. Not like I’m performing for a skeptic.

The banned list: Phrases the voice never uses. Mine includes “in today’s rapidly evolving landscape,” “it’s important to note that,” “game-changer,” “synergy,” and “let’s dive in.” Every time I catch AI using these I add them to the list.

The opinion stance: My voice has opinions. It disagrees with things. It calls out bad advice directly. A lot of AI output hedges everything - “while some may argue,” “it depends on your situation.” My voice doesn’t do that. If I think something’s wrong, I say so.

The structure patterns: I tend to start with the counter-intuitive point, then explain it. I use short punchy sentences for emphasis after a longer setup sentence. My paragraphs rarely exceed 3-4 sentences.

The Persona Injection Method

Once you have the voice guide, you inject it into every AI prompt. Every single one. Not a summary - the full thing.

I know that sounds like a lot of tokens. It is. I don’t care. The difference in output quality between a prompt with and without the full voice guide is significant. Don’t trim it for efficiency.

My prompts start with: “You are writing as [name]. Here is the complete voice guide for this writer: [500 word guide]. Now, following this voice exactly…”

The AI will still slip into genericism sometimes - especially in transitions and conclusions, which are where bad habits live. That’s what the editing pass is for.

Capturing Voice from Existing Content

If you have a lot of existing content, you can use AI to help you extract your own patterns. Drop 10-15 of your best pieces into Claude and ask:

“Analyze these pieces and identify: the average sentence length, the most common structural patterns, the phrases and transitions this writer uses repeatedly, the topics they take strong opinions on, and any phrases or constructions they seem to avoid. Then write a 400-word voice guide based on what you’ve found.”

The output won’t be perfect but it will catch patterns you didn’t realize you had. I did this and discovered I almost always open a new section with a contrarian statement. Didn’t know that until the AI pointed it out.

Real-Time Voice Checks

I use a simple Claude prompt as a final check on anything AI-generated before it goes live:

“Read this piece of content. On a scale of 1-10, how well does it match this voice guide? [voice guide]. Flag specific sentences that sound generic, over-hedged, or out of character. Suggest rewrites for those sentences.”

It catches things I miss on a read-through because I’m too close to it.

The Drift Problem

Here’s something I’ve noticed over time: AI voice drift. If you use slightly different prompts across different batches of content, the voice subtly shifts. Over 50 posts you end up with an inconsistency that readers feel even if they can’t name it.

The fix is version control for your prompts. I keep my master voice guide and my master drafting prompt in a dedicated document. When I update it, I log what changed and why. The goal is to treat the prompt like product code - versioned, intentional, not casually edited.

When AI Voice Is Fine and When It Isn’t

AI-approximated voice works well for: educational content, how-to posts, informational SEO content, FAQ pages, product descriptions.

It works less well for: founder stories, personal essays, content that requires genuine emotional vulnerability, anything where the reader is choosing to follow you specifically as a person.

For the second category, AI can be a structural scaffold. The actual words should be yours.

The test I use: if a reader found out AI wrote this, would they feel deceived? If yes, write it yourself. If they’d just shrug and say “yeah that makes sense given the volume,” you’re in the safe zone.

Know which zone you’re operating in. That’s the whole game.