The failure point in a long AI-generated video is rarely a single spectacularly bad frame. It is the succession of small disagreements between shots: the jacket changes, a doorway moves, the camera reverses screen direction or a table has different objects after a cut. A production shot bible makes those decisions explicit before generation starts and leaves the editor with a consistent set of approved assets.
Build the story into beats
Break a script into action units that can be shown. Every beat needs a narrative job: establish geography, introduce a person, demonstrate a mechanism, reveal a consequence or let the audience react. Identify moments that require a literal visual record versus symbolic illustration. For a historical or scientific film, distinguish reconstructed imagery from evidence sourced from archives; a generated visual may be expressive, but it cannot authenticate the event depicted.
A useful shot card stores a canonical ID, script line range, narration timecode, planned seconds, target aspect ratio, subject, setting, costume, camera, motion, source reference IDs, generation version and acceptance criteria. The card should also state what may change. ‘Same courier and red jacket throughout’ is a stronger constraint than ‘consistent look’, but still cannot force a model to obey it without suitable reference and editing controls.
Lock references before motion
For image-to-video, create legally cleared, approved opening stills. Use a front, side and prop-reference sheet for critical recurring subjects only where the selected tool supports such guidance. Some models accept several references; other versions explicitly do not support a character-reference feature. Google documented up to three Veo 3.1 references, while Luma’s Ray3.14 guide explicitly lists character reference as unsupported. Don’t call these capabilities generic industry standards.
In each scene, prioritize a single readable movement. The character may cross a corridor; the camera might dolly left. Asking both to reverse, change outfit and switch daylight to night in the same six seconds creates many opportunities for failure. Keep the last half-second of a shot compositionally usable so the editor has a clean transition point.
Review continuity at joins
Import accepted renders into a timeline in story order. Check the eyeline, screen direction, prop location, weather, shadow angle and movement speed across every cut. A hard cut can hide changes in style better than a forced synthetic morph; conversely, matching movement can mask minor color differences. Audio bridges and room tone can make separate short shots feel like a continuous scene without claiming they were generated together.
Record rejects with reason codes such as IDENTITY_DRIFT, PROP_ERROR, CAMERA_REVERSAL, TEXT_GARBLED or PHYSICS_FAIL. Do not regenerate acceptable clips just because newer model versions exist: keep the original selects and make improvements incrementally.
Handoff discipline
The final project directory should contain approved stills, licensed source materials, generated originals, accepted shots, captions, music rights, model IDs and an edit decision list. A shot bible is a creative control tool, not evidence that the underlying renderer or its outputs were independently tested.
Worked production example
A documentary scene calls for a courier crossing a street, looking back and entering a building. Separate the action into an establishing shot, a crossing shot and a door-entry detail. Lock coat color, bag, travel direction and the doorway design in the bible. A generator may make each shot individually attractive but reverse the courier’s direction between cuts. During assembly, check geography with a simple map and fix the transition using a motivated insert if needed. Keep the character’s approved references distinct from the final clips so isolated failures can be rerendered.