Why not both? What do you see wrong with having gpt spit out a complicated ffmpeg filter chain and explaining every step of the output, which you can then look up in the docs and unit test before implementing? I find verifying the gpt output is still quicker than consulting giant wall of text man pages