Film Crew

Film Crew acts as a producer for AI video generation: instead of writing one prompt and hoping, it runs an idea past a sequence of crew roles, each with one job, then hands over a shot list and one prompt per shot that a video model can actually execute. It targets the specific, fixable reasons AI video clips fail - two actions crammed into one clip, no camera instruction, a subject described differently shot to shot, physics the model can't render in five seconds, or a prompt that describes a feeling instead of a frame.

It has three modes. Plan mode takes an idea, script, or product and asks at most three questions before defaulting the rest (9:16 vertical, 15 seconds, model-agnostic prompts, text-to-video unless a real product or person needs image-to-video for continuity). It then runs the idea through a crew in order: Director (logline, beat sheet, hook), Production Designer (a continuity bible of locked descriptors for every recurring character, prop, and location), Director of Photography (shot size, lens, angle, height, one camera move per shot), Gaffer (time of day, key light direction, color temperature, practicals), Editor (shot durations that fit the model's clip length, cut types, first-frame hook), Sound (only if the target model generates audio or it will be added in edit), and a Script Supervisor with veto power over continuity, feasibility, and cliche ("unslop") issues - any shot it flags gets rewritten before output.

Each shot's prompt is assembled in a fixed order - shot size and lens and camera move, then the subject verbatim from the continuity bible doing one visible action, then location verbatim from the bible, then lighting and style - unless the target model's own adapter specifies a different order, in which case the adapter wins. Continuity descriptors are pasted verbatim, never paraphrased, and on-screen text or logos are excluded unless the model is known to render text.

Fix mode takes a prompt that keeps failing, names the specific problems (most damaging first, max five), and rewrites it using the same prompt order, splitting it into two shots if it was trying to do two actions at once. Review mode takes a description or shared clip of a bad generation, classifies the failure, and gives exactly one change to make before the next reroll - changing five things at once means learning nothing from the attempt.

Use it to plan an AI video - an ad, a short, a music video intro - before generating any clips; to diagnose and rewrite a text-to-video or image-to-video prompt that keeps failing; or to decide the single most useful change to make after a bad generation before rerolling.

Do not use it to actually generate, edit, or render video - it produces planning documents and prompts only, handed off to whatever video model or tool does the generation.

Input is a video idea, script, product, a failing prompt, or a description/share of a bad generated clip. Output in Plan mode is a SHOT_LIST.md containing short crew sections, a shot table, one copy-paste prompt block per shot (plus a negative prompt where supported), a reroll plan naming the most likely failure per shot and what to change if it happens, and a closing line naming the highest-risk shot and why. Fix mode outputs a before-and-after prompt rewrite; Review mode outputs a failure classification and one recommended change.

Model-agnostic by default - it writes a generic prompt structure and adapts it to a specific target video model's known adapter and clip-length limits when one is named, but it does not call or integrate with any video generation API itself.

Anyone planning AI-generated video content - ads, shorts, music video intros - who wants a structured shot list and executable prompts instead of ad hoc single prompts, plus a systematic way to diagnose and fix failed generations.