A model drifts when the words are loose. So lock the words first, and lock the frame first. The chain we use: Idea: what the shot is, in one sentence. Keyframe: the first frame, locked. Camera, subject, light. Everything else is built off this frame. Beats: the motion, broken into shots. One beat, one generation. One generation per shot: five seconds at a time, no more. Short generations stay clean. Continuity: same camera, same subject, same light from shot to shot. Assembly: shots in order, then voice and sound last. The words are the method. Write them down verbatim before you generate, and keep them the same across every shot in the chain. Change one thing at a time when you iterate, or you will not know what worked. This is how the 15 methods in the free library are built. Each one is a locked version of this chain, tuned for one kind of output.