The Scene Bible method — consistency across every AI shot

The single biggest tell of AI video is not bad motion — it is INCONSISTENCY. The character's dress changes between shots. The car gains a spoiler. The license plate reads three different things. Viewers cannot name it, but they feel it, and they scroll. The fix is structural, not luck: a SCENE BIBLE. Stage 1 — Generate one master still. One wide keyframe that contains everything canonical: the character (outfit, hair, accessories), the product/vehicle, the location, the lighting. Iterate on THIS image until every detail is right. Stills are cheap; video is not. Fix problems here, never downstream. Stage 2 — Write the bible as text. From the approved master, write a short locked description of every canonical element, e.g. "backless halter black mini dress, strappy stiletto sandals, long dark wavy hair, thin ring" — exact garment construction, not "a black dress." This text goes VERBATIM into every subsequent prompt. Models drift toward generic defaults; the bible text is the anchor that stops them. Stage 3 — Derive angles from the master, never from scratch. Every new keyframe (new angle, new framing, new moment) is generated as an EDIT of the master via an image-reference model (Nano Banana class): "same scene, same character, new camera position: low front three-quarter." The model inherits what it can see in the reference; the bible text covers what it cannot. Stage 4 — Animate keyframes, not prompts. Video generation always starts from an approved keyframe (start_image), never from text alone. Text-to-video re-rolls the dice on every canonical detail; image-to-video only has to move what already exists. Rule of thumb: if a detail matters, it must exist in pixels (the reference) or in locked words (the bible). Anything left implicit WILL drift.