AITuber
Testing Human-Like Conversation and Persona Drift with AITuber Amane
The author tests AITuber Amane's persona in Claude Sonnet 5.5 across four experiments on conversational style, topic changes, resistance to pressure, and comment-response selection. Persona instructions alone retained assistant-style habits, while additional rules shortened replies and brief stage notes at topic boundaries reduced observed character drift. The tests also exposed tensions between kindness and correction, and between responding to every comment and keeping replies manageable for a stream. Each condition was tested only twice, with evaluation relying on another model instance.

A kind persona that avoids correcting mistakes needs a closer look. I would test consistency of voice separately from the ability to disagree when needed.