A continuity technique addressing the cross-shot half of Continuity Is the Creator's Responsibility (Not the Model's): lock character/car/object blocking using an external position reference — a position-log screenshot, a hand-drawn diagram, or colored position markers — uploaded to Claude so it can be referenced when drafting each new shot's prompt. Prevents subjects from 'teleporting' between shots that are otherwise generated independently.
The position-reference image doesn't have to be generated fresh — it can be pulled directly from a video already generated for the scene. Screenshot a frame from an approved video (e.g. one showing a car correctly parked) and reuse that frame as the fixed spatial reference — the "position log" — for every later shot in that location.
This pairs with Video-Sequence Generation for Frame-to-Frame Consistency: because a generated video's frames are already internally consistent, a frame pulled from it makes a more reliable position reference than a separately-generated still.
After a video generation nails an object's placement in a location, pause and screenshot that frame, then reuse it as the interior/exterior positional reference for all subsequent generations in that scene.
When purely textual position language fails to control blocking — cars or actors drifting or 'teleporting' between shots — the fix is a visual reference rather than a better-worded prompt. Two low-tech variants are used alongside screenshotting an approved video frame as the position log:
Both work because spatial/visual instruction outperforms verbal description for blocking control — the model matches a drawn mark or line far more reliably than it parses 'park facing north, three car-lengths from the curb.'
Из тем: Unsorted, Holding Continuity Across Independently Generated Shots