AI Video Generation Prompting

There is a difference between the clips people post as curiosities and the clips people use in paid work. The second kind comes from treating the prompt as a shot sheet rather than a description of a vibe, and from generating for the cut rather than for the clip. SPEC: the seven components every prompt carries — subject, action, camera, movement, lens and depth, lighting, environment — in a deliberate order, because front-loaded terms carry more weight. Plus what to leave out: the stacked quality words inherited from early image-model habits mostly dilute a prompt now, and one style anchor beats six adjectives. SEED: text alone cannot hold a face, a product or a brand across shots, so image-to-video is the default for anything that must stay consistent rather than a fallback. Build the reference set before generating any video, use first and last frame conditioning to make cuts invisible, and fix the seed while iterating so you can tell which edit caused which change. SWEEP: every generation is a sample from a distribution, so regenerating an identical prompt and hoping is the fastest way to waste a budget. Generate four deliberate variations instead of four retries, budget against a realistic one-in-four-to-six hit rate, and kill a shot after three well-specified failures because it is fighting a model limitation, not a prompt problem. STITCH: cut early because clips degrade towards their end, vary shot size between cuts, and commit to sound design — silent AI video looks synthetic to everyone, and the same footage with room tone and foley passes without comment. Includes the disclosure, likeness and model-licence rules that decide whether the output is sellable at all.