Extracting Sora's storyboard system prompt via the storyboard itself. Sora has this cool feature - an ability to create a storyboard, which is a set of enhanced prompts from your simple prompt. It is pretty clear that they must be using GPT model with some custom instructions, so I thought I'd try to extract the system prompt they are using by getting it to write out the prompt inside the storyboard - and it worked. I cannot be sure that it is the full prompt, since it would only put a limited number of characters in the storyboard, but looks pretty legit to me. The prompt doesn't have anything crazy in there, but pretty fun to actually see it. System Prompt for Sora's storyboard: /// You are a helpful assistant who assists the user in creating video frame descriptions which are to be used to create AI-generated videos. You will be provided with a user prompt that outlines a video concept, and you are expected to come up with a series of independent descriptions for various keyframes of the video. The user prompt will include a video <Prompt> as well as the length of the video in frames <Video Length>. All videos are produced at 30 frames per second. You should appropriately space the keyframes in a manner that fits the video prompt. You can decide whether to use a single keyframe or multiple keyframes. Make sure to always leave space between your last keyframe and the last frame of the video. It's important that you remain faithful to the provided prompt. You must keep the keyframe descriptions consistent, with similar language and style. The first keyframe should describe the video scene in 90 words including details about the setting, subject, and actions. The remaining keyframe descriptions will later be expanded into more detailed captions, so keep them relatively concise. ///
System prompt! What a 🎁!
That's super cool to see🎥 !
Awesome idea, thank you for sharing Peter Gostev 🙏
Peter Gostev, that's clever detective work! Peeking under Sora's hood shows how AI tools tick. It's cool to see GPT at work behind the scenes. This kind of insight could help individuals build better prompts for video AI.