Write the short voice persona for an assistant, from its own system prompt.

A speech-to-speech model that delegates reads two prompts: the full system prompt goes to the model doing the thinking, and is consulted only on the turns handed to it. The voice model answers the rest itself — about four turns in five on real calls — and reads this instead. Without it those turns come out as a generic chatbot.

Written by default with the model the voice session delegates to, so the persona is written by the same model that will be answering beside it.