Footage generation, run like a film set.

Before anything generates, the project gets two locked documents. A settings file hard-locks the machine decisions: aspect ratio, resolution, which model handles what, texture and audio rules. Nothing drifts between sessions because nothing is allowed to.

The bible holds the creative truth. Its core is a section we call the Visual DNA: a frozen style prefix and a frozen constraints core that every single generation must open and close with, verbatim. If the Visual DNA has not been authored, the generate command refuses to run.

Every character, environment, prop, and key asset gets a reference sheet before any shot exists. Neutral grey backdrop, even light, multiple views, and an ID in the project's element registry.

Once approved, an element is reusable everywhere. Any later prompt binds that exact character or prop by ID, with no re-upload and no re-describing. Elements survive the project too, which is what makes cross-project reuse a one-line instruction instead of an archaeology dig.

The hard rule of the whole system: never text-to-video. Every video starts from a frame we already approved, or from locked references. A finished generation can itself become a reference, which lets a look or an identity hold across an entire shot series.

Shot planning draws on a 186-entry bank of real film language: shot sizes, angles, camera moves, lighting setups, director recipes. Built from actual cinematography sources, not from what a model thinks a movie looks like. Heavy batches fan out to parallel agents, one per shot.

Generation is cheap. Judgment is not. Every batch lands in a review gallery with thumbnails, approve and reject controls, and keyboard shortcuts, and a human moves takes from raw to selects to final.

Approved takes get promoted into an editor-ready footage folder, grouped by section. The editor never sees the noise, only the cut candidates.

Every output carries a sidecar record: the exact prompt, the references it used, the model, the settings, the cost, and an append-only history of every revision. Any frame in any delivery can be traced back to exactly how it was made.

That record is also the archive. When a client asks for one more shot six weeks later, the system reads the sidecars and the bible and picks up where it left off.

Version one was a pile of scripts we wrote ourselves: hand-rolled API clients for every model, a batch processor, custom review pages. It worked, and then it collapsed under its own weight. Context lost between sessions, prompts lost on rework, versioning by duplicating folders, ninety-eight loose scripts, and four different places all claiming to be the memory.

Version two was built in one long session after a proper autopsy of version one. Every failure became a rule: one always-current status doc, prompts live only in sidecars, one bible per project, one unified generation engine instead of a client per model. The system got simpler by getting stricter.