How to Train AI on Your Brand Voice

Build an AI writing voice indistinguishable from your own

You’ve been prompting AI to “write in a friendly, conversational tone.” That’s not your brand voice. That’s everyone’s brand voice.

You ask AI to write a LinkedIn post in your voice. It produces something polished, competent, and completely indistinguishable from every other AI-written LinkedIn post. It sounds like an expert. Not like you.

The problem: “Friendly and conversational”

Everyone tells AI to be “friendly, conversational, professional.” That describes 90% of the internet.

AI is trained on the statistical average of internet writing. It predicts what comes next based on patterns from millions of documents. Most of that writing is middle-of-the-road. It is professional but forgettable, optimized for search engines, not human connection.

So when you say “write in a friendly tone,” you get the average of all “friendly” writing on the internet.

Here’s what that looks like:

Generic AI output:

“Artificial intelligence has changed the way businesses operate. By using AI tools, you can streamline your workflows and achieve new levels of productivity. The key is to implement these solutions strategically.”

Voice-trained output:

“AI tools promise to save you time. Then you spend three hours watching tutorials and another two trying to connect your apps. Here’s what actually works: Start with one task that takes you 30 minutes every day. Automate just that. Nothing else.”

Same topic. Completely different voice.

The second one sounds like a person. That’s the goal.

You can tell AI-written content not because it’s bad, but because it’s generic. It could have been written by anyone. It has no fingerprints.

Your brand voice is the one thing that should be impossible to replicate with a lazy prompt. That’s what makes you unpromptable.

But here’s the real problem:

You’ve never documented what your voice actually is. You know it intuitively. You recognize when something sounds like you. But you’ve never written down the patterns that make it work.

So you default to adjectives. “Be witty. Be concise. Be bold.” And AI gives you its version of witty. Which isn’t yours.

Why “write like me” doesn’t work

Adjective-based prompting is the core issue.

When you tell AI to “be witty,” you’re asking it to match a subjective standard. AI doesn’t know what your version of witty looks like. It knows the statistical average of “witty” across millions of documents. Your witty might be dry one-liners. AI’s witty is whatever gets the most engagement across the internet.

Your voice isn’t a list of adjectives. It’s three things:

1. Style rules – The measurable patterns: sentence length, vocabulary choices, structural decisions

2. Writing examples – Concrete samples AI can study: your best posts, emails, scripts

3. Constraints – What you never do: jargon you avoid, topics you skip, tone boundaries

Think of it like a recipe, not a restaurant review.

Saying “write like me” is like telling someone to “cook like grandma” based on how the food tasted. They’ll make something vaguely Italian. But give them grandma’s actual recipe. Her specific measurements, her technique for browning the garlic, her rule about never adding oregano and you get Sunday dinner.

The gap isn’t that AI can’t learn your voice. It’s that you’ve never documented it in a way AI can use.

That’s what this process fixes.

The brand voice training process (3 steps)

The fix isn’t more adjectives. It’s a system. Three steps to train AI on your actual voice, the documented version, not the adjective version.

Step 1: Document your patterns (The Voice Audit)

Pull 5-10 pieces of your best writing. The stuff that sounds like you.

Choose variety:

• A LinkedIn post that got engagement,

• A newsletter readers responded to,

• An email that was converted.

Pick content where your voice comes through clearly.

Feed them to AI with this prompt:

Analyze these writing samples and create a voice profile covering:

1. SENTENCE STRUCTURE - Average length, variation patterns, use of fragments

2. VOCABULARY & TONE - Formality level, common phrases, frequently used words

3. STRUCTURAL PATTERNS - How I open, transition, and close pieces

4. DISTINCTIVE QUIRKS - Signature phrases, rhetorical devices, formatting choices

5. CONSTRAINTS - Jargon I never use, tones I avoid

Don’t summarize with adjectives. Give me specific, measurable patterns with examples from the text.

What you get back is your “Voice DNA” document.

Not “friendly and conversational.” Instead: “average sentence length of 12 words, opens 80% of pieces with a question or contradiction, never uses ‘leverage’ or ‘synergy,’ transitions with ‘here’s the thing’ or ‘look.’”

Specific. Trainable. Save this document. It becomes your reference for every AI conversation going forward.

Here’s what makes this different from what you’ve been doing:

Instead of telling AI who you are at the start of every chat, you hand it a document that describes who you are in terms it can actually replicate. You stop being a vague creative direction. You become a spec.

Step 2: Create annotated examples (The Training Set)

Take 3 of your best pieces and annotate them.

Mark WHY each section work in your voice. This is the difference between giving AI your writing and teaching AI your writing.

Most people skip this step. They paste their writing into AI and say “write like this.” That’s like handing someone a painting and saying “paint like this” without explaining brush technique, color mixing, or composition choices. You get a copy that misses the point.

Annotations make the implicit explicit.

Here’s what they look like:

• “Notice I used a metaphor from cooking here. I always ground abstract concepts in everyday analogies”

• “This paragraph is just two sentences. I keep impact paragraphs short.”

• “I said ‘look’ before making a direct point. That’s a verbal tic I use in writing.”

• “I opened with a contradiction (promised time savings, then showed time waste). I do this in 70% of my intros.”

You’re showing AI the decision-making process behind what you wrote.

Once annotated, feed them to AI:

I’m going to share a piece of my writing with annotations explaining WHY I made specific choices. Study these patterns, don’t just copy the words, understand the underlying decisions.

[Paste annotated example]

Now write a 200-word piece on [topic] following these same patterns.

After writing, explain which of my patterns you applied and where.

The “explain which patterns you applied“ part is critical. It forces AI to think about patterns, not just mimic words. You’ll see AI start referencing your actual decisions “I used a short paragraph here because you tend to isolate key points” instead of producing generic structure.

That’s when the output shifts from “AI wrote this” to “this sounds like me.”

Step 3: The testing loop (Calibration)

Generate a piece. Read it. Score it on a “sounds like me” scale of 1-10.

The benchmark: aim for 70%.

That means output you’d publish with light editing. Not something you’d completely rewrite. You’re not looking for perfect. You’re looking for “I can fix the last 30% faster than writing from scratch.”

When it misses, don’t say “try again.” Be specific:

Rate: 4/10. The structure is right but the tone is too formal.

Specifically:

- I never say “it’s important to”. I’d say “here’s what matters”

- My paragraphs are shorter, break that second paragraph after the question

- I use “you” more than “one” or “we”

Try again with these corrections.

After 3-5 rounds of specific feedback, AI starts nailing your patterns. Each round gets closer because you’re giving it concrete corrections, not vibes.

The key is being diagnostic. Not “this doesn’t sound like me.” But “you used passive voice in paragraph three and I never do that.” Not “the tone is wrong.” But “I wouldn’t say ‘it’s worth noting. I’d say ‘here’s what matters.’”

Once you hit 70% consistently, save the conversation as your “voice template.” Most AI tools let you save system prompts or conversation contexts. Use that feature. You don’t want to retrain from scratch every session.

But sometimes the output still feels off. Here’s how to fix it.

When it still sounds like AI (Troubleshooting)

Still too formal?

Don’t say “be more casual.” Say: “Re-read sample #3. Match that level of casualness, not what you think ‘casual’ means.”

AI keeps using words you hate?

Be explicit. List the exact banned words.

“Never use ‘leverage,’ ‘utilize,’ ‘innovative,’ or ‘streamline.’ Replace with plain English: use, new, improve.”

Structure doesn’t match?

Replace vague instructions with specific patterns: “Short paragraphs (3-4 lines max). Start 80% of sentences with ‘You’ or action verbs.”

Still need heavy editing?

Add 2-3 more samples of the specific content type you’re creating. LinkedIn posts train LinkedIn voice. Newsletter intros train newsletter voice. Don’t cross-train with mixed formats. AI needs enough examples of that specific context to match your patterns in it.

Now for the final test.

The “Unpromptable” Test

You’ve done the work. You have a Voice DNA document, annotated examples, and a calibrated voice template. Now run this test:

Show the AI output to someone who reads your work regularly. A colleague, a subscriber, a friend who knows your writing.

Ask: “Did I write this or did AI?”

If they can’t tell, you’ve succeeded.

But here’s the deeper win. This process doesn’t just make AI output better. It forces you to understand what actually makes your writing yours. Most people have never done that exercise. They know their voice intuitively but have never broken it down into patterns. That’s why their AI output sounds like everyone else’s, they can’t teach what they’ve never articulated.

The paradox: by training AI to write like you, you become more unpromptable. Replicating your voice now requires your Voice DNA, your annotated examples, your constraints. That’s not something someone can prompt their way into with “write like a thought leader in the AI space.”

Remember “friendly and conversational” from the beginning? That’s everyone’s voice.

You now have the one voice that can’t be prompted into existence by someone who doesn’t know you.

That’s the real payoff. Not just better AI output. A deeper understanding of what makes your writing yours and a system to scale it without losing it.

Key takeaways

1. Stop describing your voice with adjectives, document it with patterns (style rules, examples, constraints). Adjectives are subjective. Patterns are trainable.

2. Use the Voice Analysis Prompt to generate your Voice DNA, a concrete reference doc, not vibes. This becomes your foundation for all AI conversations.

3. Annotated examples teach AI why your writing works, not just what it looks like. Show AI your decision-making process, not just your finished work.

4. The testing loop with specific feedback is where the magic happens, aim for the 70% benchmark. You’re not looking for perfect. You’re looking for “I can edit this faster than writing from scratch.”

5. Your brand voice is your competitive moat, train AI to amplify it, not replace it. Documenting your voice makes you more unpromptable, not less.

6. Run the “Unpromptable” test, show your AI output to someone who knows your work. If they can’t tell the difference, you’ve succeeded.

Now go document your voice. Make yourself unpromptable.

---

PS. From Dheeraj Sharma

Want to take this further? I wrote a deeper guide on setting this up with Claude Projects for persistent voice memory across sessions. Check it out on GenAI Unplugged.

And if you’re a Substack writer, check out my tool “SubflowAI” - it repurposes your existing articles and bulk schedules months of Notes in less than 10 minutes.