Written camera direction is approximate. Ask for "a whip pan down past the skid and a near-vertical hold" and you get something in that neighbourhood, take after take, at full price per take. A keyframed camera is not approximate. Seedance 2.5 and MiniMax H3 both accept a reference video, and when that video is an untextured grey blockout they follow its camera path, framing, blocking and timing almost exactly, then invent everything else from your paste.
This lesson builds the blockout. Lesson 12, Blender previz: submit and iterate, submits it and reads the takes.