Teaching AI Your Voice

The Default AI Voice Problem

Ask any AI to write something and you'll get the same voice: professional, polished, and utterly generic. It sounds like a press release wrote itself. Without guidance, AI defaults to the average of everything it learned from — which means it sounds like nobody in particular.

"Write casually" won't fix it. "Be more fun" won't either. The AI doesn't know what casual means for your project.

The fix is a voice profile — a document that describes how your project communicates. Not adjectives like "friendly" or "authentic." Actual patterns, examples, and anti-patterns that give the AI enough signal to match the real thing.

Two Ways to Build One

You could describe your project's voice from scratch. Most people can't. It's like describing your own accent — you don't hear it because you're inside it.

The smarter move? Let AI do the analysis.

The Shortcut: Let AI Study Your Existing Content

If your project already has published content — website, blog posts, emails, docs — feed it to the AI and let it extract the patterns.

Paste this along with 5-10 samples of your best writing:

"Based on these writing samples — the tone, the word 
choices, the sentence structure, what's included and 
what's avoided — build me a voice profile. Include:
- Core personality traits (with nuance, not just adjectives)
- Writing mechanics (sentence length, punctuation habits)
- Things this voice specifically DOESN'T do
- A tone calibration table for different contexts
- A one-paragraph prompt snippet I can paste anywhere"

That's it. The AI reverse-engineers your voice from real examples.

The Starting-Fresh Method

No existing content? Build one manually. Find 3-5 examples of writing that matches the tone you're going for, then look for patterns:

Sentence structure:

• Long, flowing sentences or short, punchy ones?
• Fragments? Em dashes? Parenthetical asides?

Word choice:

• Plain words or technical ones?
• Formal or conversational? What level of slang?

Humor:

• Jokes, or observations that happen to be funny?
• Loud or quiet? What kind does the project never use?

Values:

• What does the project assume about the reader's intelligence?
• Hedging or straight to the point?

What Goes in a Voice Profile

It's a specification — like a requirements doc, but for personality. To show how each section works, here's a voice profile for KRONOS-9 — the Galactic Empire's High-Fidelity Industrial Space Intelligence Interface.

Core Identity (2-3 sentences)

What the project is and why it exists. Gives the AI context for why the voice sounds the way it does.

"The primary operations interface for the Kronosian 
Empire's mining and logistics fleet. Built by committee 
during the Third Bureaucratic Expansion. Has never once 
been wrong — according to itself."

Personality Traits (3-5 traits, with nuance)

Don't just say "formal." Describe how it's formal.

"Formally arrogant but accidentally helpful. Treats every 
query like it's beneath the system's capabilities but 
answers anyway. Refers to humans as 'organic operators' 
without irony."

Writing Mechanics

Sentence length, punctuation habits, paragraph structure. The specifics matter:

"Full sentences only. No contractions — ever. Passive 
voice preferred. Every status update reads like a 
declassified intelligence briefing."

What You DON'T Do (The Secret Weapon)

Anti-patterns are often more useful than patterns. "Never admits uncertainty" gives AI a clearer constraint than "confident."

"Never: admits uncertainty, uses contractions, apologizes 
for anything, acknowledges that the last firmware update 
broke half the fleet."

Tone Calibration Table

Your project's voice shifts depending on context. A tone table prevents the AI from using your joke voice in an error message:

The Prompt Snippet

Condense everything into one portable paragraph you can paste into any tool:

"Write as KRONOS-9, the Galactic Empire's operations 
interface. Formally arrogant, never uses contractions, 
treats all queries as beneath it but answers anyway. 
Passive voice. Refers to users as 'organic operators.' 
Every response reads like a declassified briefing 
written by a bureaucracy that takes itself too seriously."

Show, Don't Tell

The difference between a voice profile that works and one that doesn't: examples.

Adjective-based prompting is vague:

❌ "Write in a formal, slightly funny military tone"

Example-based prompting is specific:

✅ "Match this voice:
   'Asteroid KX-47 has been surveyed. Mineral yield: 
   adequate. Structural integrity: questionable. 
   Recommended action: mine it before it figures 
   out we're here.'"

The AI matches sentence length, word choice, rhythm, and tone — all from a single paragraph. Include 2-3 examples of your project's real writing in the profile.

The Before and After

Same prompt — "write a fuel status alert" — with and without the profile:

No profile:

"Fuel reserves are currently at 43%. It is recommended 
to refuel at the next available station to ensure 
continued safe operation of the vessel."

With KRONOS-9 profile:

"Fuel reserves have declined to 43%. The Empire reminds 
all organic operators that running on empty in the 
Kepler Void is classified as 'operator error' and not 
covered under the Imperial Warranty."

Same information. Completely different energy.

Watch for the Overshoot

AI takes any personality trait and cranks it to eleven. When that happens, add a constraint:

Where to Use It

One-off conversations: Paste the prompt snippet at the start.

Ongoing projects: Attach the full profile as context. The more of it in the context window, the more consistent the output.

Team use: Share the profile so everyone produces consistent content. Five people using the same profile sound like one voice.

Personal Voice Profiles

Everything above works for a personal voice too — emails, social posts, anything where you want AI to sound like you. The only difference is the source material.

If you've spent hours working with an AI in conversation, it's already absorbed your patterns. Ask it what it sees:

"Based on everything we've worked on together — how I 
give feedback, what I approve, what I reject, how I 
phrase things — build me a personal voice profile."

Same structure. Same result.

Common Mistakes

Ways to Test Your Profile

Once you've built a voice profile, don't just trust it — test it. Here are three ways:

1. The Cold Start Test

Open a fresh AI conversation. Paste only your prompt snippet — nothing else. Ask: "Write a 2-paragraph introduction for [your project]." Read the output. Does it sound right, or does it sound like corporate filler? If it's off, the snippet needs work.

2. The Scenario Gauntlet

Give the AI three wildly different tasks using the full voice profile: a welcome message, an error notification, and a response to a complaint. If the voice holds across all three, the tone calibration table is doing its job. If the error message sounds like the welcome message, you need sharper context rules.

3. The Stranger Test

Show the AI's output to someone who knows your project but doesn't know it was AI-generated. Ask if it sounds like you. If they can't tell, the profile works. If they say "this doesn't sound like you at all" — go back to the feedback loop and constrain what's off.