Why Your AI Sounds Like Everyone Else

You sit down in Claude or ChatGPT and ask it to write something in your voice. Warm, witty, smart. Maybe punchy and conversational, if you want to get fancy. The AI nods (figuratively) and produces a LinkedIn post that sounds like a content marketing manager from 2019.

So you prompt harder. You add more adjectives. You tell it to "make it sound human," which is a thing we apparently say to robots now. Still nothing. The output is competent and forgettable, which is somehow worse than bad.

You blame the prompt. You blame the model. You consider blaming Mercury.

The actual problem is the source material. You haven't told the AI who you are. You've told it who half of LinkedIn already is.

How I figured this out (also: a brief flex)

Before I worked in content strategy, I spent two years at Appen as their Head of Content, on the marketing side of large language models. That's the side where teams of humans shape what AI learns from. So I know a little about training data, what makes it good, and why "more" is not the same as "better." Before Appen, I studied psychology, which gave me a different question to ask about AI than most marketers ask. Not "how do I prompt it," but "what does it actually need to produce the output I want?"

When ChatGPT became something normal people could use, I didn't approach it like a prompt engineer. I approached it like someone who knew the AI wasn't going to magically figure out who I was. I'd have to give it enough specific information to recognize me. So I built a brand voice document. A real one. Long, weird, deeply specific, and weirdly satisfying to write.

The voice held. Across platforms, across years, across model upgrades. It still holds now.

I want to share what I learned, because most of the "train AI on your brand voice" advice floating around right now is built on the wrong premise.

The clarification I'm legally obligated to make

When you upload a document to a Claude project or paste samples into ChatGPT, you're not actually training the AI. The AI itself doesn't change. What you're really doing is giving it a great reference document to look at every time it writes for you. Think of it less like teaching a student and more like handing a writer a thick briefing packet right before they sit down to draft.

This sounds like a small distinction. It's not. It changes how you should think about the document. You're not feeding a brain that grows over time. You're building a reference packet so detailed that the AI can pull out exactly what it needs every time. Which means the document has to be good, and good in a very specific way.

So what does "good" look like? Not adjectives. Read on.

Adjectives are where voice goes to die

"Warm, witty, smart." Cool. So is everyone. Every single brand voice template ever filled out has used some version of those words, and that's exactly why the AI output sounds generic. You're handing the AI a description that fits ten million other people, then acting surprised when it produces content that could belong to any of them.

Adjectives are where
voice goes to die.

Voice doesn't live in adjectives. It lives in specifics.

Actual phrases you say. Things you find genuinely funny. What you're a snob about. What you'd never reference because it doesn't feel like you. The weird contradictions that make you a real person and not a LinkedIn caricature.

When I built my own document, the answer to "what phrases do you say all the time" wasn't "I'm conversational." It was "ain't no thing but a chicken wing" when someone thanks me, and "hey hey!" at the top of every email. The answer to "what's your humor like" wasn't "dry and clever." It was that I recently lost it at a fanfic tag joking that Rhysand from ACOTAR would absolutely be in a group chat called The Inner Circle. The answer to "what are you a snob about" was romantasy, AP style, and the difference between a well-crafted social post and a content calendar filler.

That's the kind of stuff the AI can actually use. It can't guess I'd reach for a Sarah J. Maas reference. It can't know I hate em dashes with the heat of a thousand suns. It can't tell that my politeness gets sharper, not warmer, when I'm annoyed but staying professional. None of that lives in adjectives. All of it lives in specifics.

Bring your whole self (yes, even the embarrassing parts)

Here's the part that took me longest to articulate.

You are not one version of yourself when you write. You're a strategist who also belts Moana on the drive to the office. A consultant who reads romantasy at night. A designer who plays in a band on weekends. All of those accumulated details, over years, shape your instincts. Not in a literal Moana-equals-Moana way. But together they form the texture of your voice. The references you reach for. The rhythm in your sentences. The kind of joke you'd actually make.

If you only give the AI the version of you wearing your "content strategist" hat, it's working from a fraction of who you are. You can't expect a complete voice to come out of a partial input. The document has to be wide enough to include the parts of you that don't seem professionally relevant but absolutely shape how you sound.

The Moana matters. The fanfic matters. The contradictions matter most of all.

More isn't more

A caveat, because more isn't infinitely better. After a certain point, piling in more context starts to dilute the signal, especially if some of it contradicts other parts. The goal isn't to give the AI everything you've ever felt. It's to give it enough distinctive, specific detail that it can recognize you and write like you.

Twenty answers that are deeply, weirdly specific to you will train a better voice than two hundred surface-level ones. "I hate em dashes and use commas, parentheses, or two sentences instead" is worth more than fifty bullet points about your tone being "approachable." Specificity is the whole game. Specificity is also, I'll note, the thing the average brand voice template is allergic to.

Aspirational, not literal

One more reframe. The voice you're building doesn't have to be exactly how you sound on a Tuesday afternoon when you're tired and three meetings deep. It can be the sharpened, more intentional version of you. The voice you'd have on your best day, with infinite time to revise.

This is freeing. You don't have to teach AI to mimic every quirk you have. You're teaching it the direction you want to lean. Your real voice plus a little more polish. Your humor with the jokes that actually land. Your references with the ones that connect.

The document captures who you are AND who you want to sound like. Both matter.

The worksheet

I built a discovery document that asks the questions I've spent years figuring out how to ask. The ones that pull specificity out of people who keep reaching for the polished-LinkedIn answer. It's the document I wish I'd had when I started, and the one I now use with friends building their personal brands.

Fair warning: it'll take time to fill out. That's intentional. The questions don't reward speed. They reward the version of you who'd answer your best friend honestly instead of giving the version you'd put on your About page.

How to actually use this thing

Filling out the worksheet is step one. The rest is where most people stop, which is also why most people are still typing "make it sound human" into Claude six months later. Here's the workflow I use myself.

1. Test before you build. Drop the worksheet and your writing samples into a one-off Claude chat. Ask Claude to review them and learn the voice. Then give it test assignments. A blog post. A social caption. A Slack note to a colleague. An email pitching something. Different formats matter, because voice shows up differently in each.
2. Refine in the chat. When the output isn't quite right, explain why. "Too formal here." "I'd never use that phrase." "This is closer, but the rhythm's off." Note what you're correcting. Those corrections ARE the voice instruction.
3. Update the document. Take what you learned in the chat and feed it back into the worksheet. The corrections are the missing context. If you skip this step, you've trained one chat instance and left the document behind to fend for itself.
4. Build the project. Open a fresh Claude project. Upload the updated worksheet, your writing samples, and ask Claude to generate a tight voice document based on the synthesis. Three layers of context: the raw answers, the distilled voice document, and the writing evidence.
5. Write project instructions. Have Claude draft instructions for that project that tell it to apply the voice to every output, flag when it's unsure, and respect your hard no's. Otherwise the project knowledge sits there passively while Claude produces something generic.
6. Build a skill. This is the unlock. A voice skill lets you invoke your voice in ANY project, not just the one you set up. Client deck. Coding project. One-off chat. Doesn't matter. You can call your voice in whenever you need it.

The whole thing is a system, not a one-time exercise. The document is the foundation. Everything else is what you build on top so the voice actually lives in your daily work, not in a Google Drive folder you'll never open again.

The bigger point

Treat AI like a search engine and you'll get search engine outputs. Treat it like a smart collaborator who needs the whole picture to make good decisions and you'll get something that sounds like you.

The whole picture part is on you. Nobody else can write your voice document. Nobody else has been belting Moana on your specific commute. The AI's only as good as the information you give it, and most people are giving it adjectives.

Give it the specifics. Give it the contradictions. Give it the Moana.

The output will follow.

---