Building a Voice DNA File (So AI Stops Sounding Like AI)

Three sections. Five minutes. One unfair advantage.

The Briefing

Meta acquires Manus AI for $3 billion.

General-purpose AI agents that execute complex tasks independently. Meta's betting its next platform shift on autonomous AI. If you're in customer service, ad ops, or content, the automation timeline just accelerated.

Claude integrates with Notion, Canva, Figma, and Atlassian.

Claude now automates workflows directly inside the tools marketers use daily. It's competing with Zapier and Make, but with actual intelligence behind the automation.

xAI raises $20 billion from Nvidia, Cisco, and Fidelity.

Musk's AI company just raised one of the largest rounds in history. Another serious competitor entering the market means more tools, lower prices, and faster innovation for everyone.

The Build

I have a problem with AI copy. You probably do too.

It sounds like AI.

"In today's competitive landscape..." "Unlock your potential..." You know the smell. Your customers know it too. And they scroll past it like it's a billboard they've seen 400 times.

So this week I built what I'm calling a "Voice DNA" file. Writing rules, formatting standards, and (critically) a kill list of phrases that I feed to the AI before it writes anything. The banned phrases did more for copy quality than any amount of positive instruction.

The concept: AI defaults to patterns. It reaches for the same dead phrases because those phrases appear millions of times in its training data. You have to explicitly block them.

Think of it as guardrails on a bowling lane. Without them, every ball drifts into the gutter.

The file that made it work: voice-dna.md

This file contains writing rules, formatting standards, and a comprehensive banned phrase list organized by category. When loaded into a project, the AI stops defaulting to AI-speak and starts writing like a human who's done the work. Here's a section:

# Voice DNA

## Writing Rules

- Write like a sharp human, not a language model.
- Use contractions naturally (don't, can't, won't).
- Short paragraphs. 1-3 sentences max.
- Get to the point. No throat-clearing, no preamble.
- If making a claim, be specific. Use numbers, names,
  concrete details.
- Use physical verbs: "sanded down" not "improved,"
  "bolted on" not "added," "stripped back" not "simplified."
- Humor comes from specificity, not from jokes.

## BANNED PHRASES (never use these, ever)

### Dead AI Language
- "In today's [anything]..."
- "Delve" / "Dive into" / "Unpack"
- "Harness" / "Leverage" / "Utilize"
- "Game-changer" / "Cutting-edge"
- "Straightforward"

### Dead Transitions
- "Furthermore" / "Additionally" / "Moreover"
- "Moving forward" / "At the end of the day"

### AI Cringe
- "Supercharge" / "Unlock" / "Future-proof"
- "10x your productivity"

### The Big One (FATAL)
- "This isn't X. This is Y." and ALL variations.
- ANY sentence that negates one framing then asserts
  a corrected one. If even ONE appears, the output fails.

Want the full file? Download voice-dna.md from the resource folder

The AI went from sounding like a LinkedIn influencer to sounding like a human who's actually done the work.

What worked: The banned phrase list. AI gravitates toward certain phrases the way a drunk gravitates toward a lamppost. You have to explicitly block them. The "FATAL" category was the biggest win. That "This isn't X. This is Y." pattern is AI's favorite crutch, and it makes everything sound like a TED talk from 2018.

What didn't work: I only gave rules about what TO do. The AI still snuck in dead language. Adding the banned list was the turning point. Positive instructions set the direction. Negative constraints enforce it. You need both.

The Play

Build your own banned phrase list. Takes 15 minutes.

Open your last 5 pieces of AI-generated content. Highlight every phrase that makes you cringe, every transition that sounds robotic, every word no real human would say out loud. Write them down. Add them to your AI instructions with "NEVER use these phrases."

Then add 3-5 writing samples that show your actual voice. The combination of "here's what I sound like" plus "here's what I absolutely don't sound like" is more powerful than any writing prompt I've tested. Your AI output goes from 60% usable to 90%+.

Try this today. The constraint creates the quality. Tell the AI what NOT to say, and what's left starts sounding like you.