Built an AI writing system that learns an author's voice from their 100 most recent posts. Not "write like X" via prompt engineering — actual style extraction: vocabulary distribution, sentence length entropy, profanity scoring, emoji patterns, structural templates.
Training takes 5 minutes per author. Output: posts so close to original style that GPT 5.5 (in adversarial testing) couldn't reliably classify which posts were AI-generated vs human-written.
Built for Telegram channel admins burning out from daily posting, but the technique generalizes to any author with sufficient post history.
Anyone here tackled the "voice replication" problem differently? Curious how others approach style extraction beyond simple fine-tuning.
Top comments (0)