Persona Distill: chat history in, AI persona out
Most AI personas are written from scratch. Startover's persona distill works the other direction — you feed it a real chat history and it extracts a working persona across six fidelity dimensions: personality, language fingerprint, emotional defaults, values, interests, and relational dynamic.
What persona distill is for
For anyone who wants a recognizable voice instead of a generic chatbot impression. For people exploring how chat-based persona extraction works in practice — the gap between "pretend to be X" and "actually sound like the X my friend knew" is what distillation closes.
How Startover distills a persona
Paste the chat (WeChat, iMessage, Line, Telegram — text or screenshots). The distill pipeline reads ~500+ messages and extracts the six dimensions in roughly 30 seconds. The result is a persona you can talk to in a sandbox, with the same patterns the source person actually used — including how they avoid things, how they argue, how they end a conversation.
Try the distill
Free during beta. Source chat is deleted after 30 days. Personas live only in your account.
Common questions
What is persona distillation?
Extracting a working persona from real messages a person sent — language fingerprint, emotional defaults, the way they actually respond — instead of writing a prompt that approximates them. Closer to knowledge distillation than to a writing exercise.
How is this different from prompting GPT to pretend to be someone?
A generic "pretend to be X" prompt produces a flat impression of public X. Startover's distill works from private messages, so the persona reflects who the person was with you — not their public-facing version.
What dimensions does the distill capture?
Six: personality (how they think), language fingerprint (how they write), emotional defaults (what they go to under stress), values (what they protect), interests (what they pull conversations toward), relational dynamic (how they were with you specifically).
How much source material does the distill need?
500+ messages usually gets recognizable on most dimensions. Less works but produces a thinner sketch. For fictional characters or deceased loved ones with little written record, the Baimiao flow uses 14 questions instead.