“Anthropic's 'character training' for its Claude models involves the model generating user queries relevant to a specific trait, generating responses, and then ranking those responses based on the desired character traits, a process similar to Constitutional AI.”