Sora2's system prompt, from what I can extract: -Instructs GPT to generate metadata for the video first -Then instructs it to generate captions for the video -Then asks it to generate the "audio you imagine" From what I can tell it seems like there is an orchestrating LLM that drafts an audio description/timeline the synthesizer then realizes(Conor)