We rebuilt Mirror's voice calibration system from scratch this month.
The problem: Generic AI outputs sound like generic AI. Even with detailed prompts, ChatGPT and Claude generate content that feels... off. Too formal. Too verbose. Too polished.
Our solution: A three-layer voice kernel that enforces style constraints at the token level.
Layer 1: Structural constraints (no emojis, max sentence length, paragraph breaks) Layer 2: Lexical preferences (preferred terminology, forbidden phrases) Layer 3: Narrative patterns (story arcs, argument structures)
The result? Content that sounds like you wrote it—because it follows the same rules you do when you write.
We're still tuning it. But early users say it's the difference between "AI-generated content" and "content that happens to be generated by AI."