Agree with @Christian McIlwain, @Jeremy Rivera . I've had the best success in taking the transcript, but then also feeding in some manual notes (takeaways) and even the luma page that stores the main items we'd talk about in promoting the event. It helps to direct the model towards what you actually want to see as output.
Secondly, using user-assistant pairs is very helpful too (not as complicated as it sounds!)
And thirdly, given you mentioned video/podcast, lately I've been using something to return for me the key timestamps from a podcast / video, which lets me or an editor snip up the important moment much more easily.
Happy to share a workflow with you all that's been working for me!