How AI can extract your brand voice from three of your past posts

Generic ChatGPT output is the most obvious tell of an AI-written post. "Unlock your potential" — every reader's eyes glaze over.

Solution: give the AI a voice profile — a structured description of how *you* write. We built this into Virelta because we use it ourselves daily.

Here's the practical walkthrough.

Step 1 — Pick your three best posts

Not your most viral. Your most *you*. The ones friends read and say "yeah, that sounds like you."

Why three: one sample teaches almost nothing. Three is the minimum for the model to triangulate patterns vs one-off quirks.

Step 2 — Paste them in

In Virelta: Brand voices → New voice → "Extract voice from examples" button. The model returns:

• A short name (you can rename)
• A 2-3 sentence description of who you are and who you write for
• 5-8 tone rules as bullets
• 3-15 banned words (things you'd never write)

Step 3 — Read it. Fix what's wrong.

The model gets ~70% right on the first try. You'll spot rules that don't match you. Edit them.

Common adjustments: - It says you use emojis — actually you hate them. Add a banned-emoji rule. - It misses your habit of opening with a specific number. Add it explicitly. - It captures your structure but misses your *attitude*. Add a tone rule like "always assume the reader is smart and busy."

Step 4 — Use it

When you create a new project, pick the voice from the dropdown. Every one of the 20 platform outputs will be generated with that voice in the system prompt.

The output will sound like you. Not 100%. But 80-90%, and editable in 30 seconds.

Why this beats "fine-tuning"

Fine-tuning an LLM requires hundreds of examples and a paid pipeline. It locks you into one model. It breaks if the model is upgraded.

A voice profile is a few KB of plain text. It works with any model. You can edit it. You can have multiple voices for different contexts (your LinkedIn voice ≠ your TikTok voice).

For 95% of creators, voice profiles are the right answer. Fine-tuning is overkill until you have a writing team and consistency requirements that justify the operational cost.

What we'd love feedback on

The extraction prompt is in lib/ai/router → it's a single LLM call with a strict JSON output spec. If you've extracted a voice and the output felt off, hit reply on any of our emails and send the input samples. We use them to improve the prompt.