I'd reckon you could get very far relying on GPT-4 to create the appropriate ffmpeg calls based on your precise needs, it has the documentation in its training corpus and is great for this kind of thing.