How we are building continuity into AI video

A clip can look impressive while still failing the sequence. The coat changes, a prop moves to the wrong hand, the camera crosses the line, or the next action starts before the previous one has finished. Each result can be locally plausible and globally wrong.

Our first design decision was to keep Project, Scene, Shot, Character, Location, Object, and Generation as separate concepts. That structure gives the agent a real workspace to reason over and gives the creator specific things to review and change.

Director keeps project-local references for the people, places, and objects that should remain recognizable. Each reference has a stable name, a selected image, and a history of generation attempts. Selecting a new result changes the pointer without erasing how the project arrived there.

This is more useful than repeating a character description inside every prompt. The name connects the creative conversation, the storyboard, and the generation inputs to the same visual anchor.

We found that aggressive shot splitting creates its own continuity tax. When a coherent scene fits inside one provider run, one shot can carry timed camera changes and complete the dramatic beat without an unnecessary boundary.

When a split is necessary, every shot needs a start state, an achievable action, and an intended end state. Duration is not only a number. It is a constraint on whether the action can honestly finish before the clip ends.

For a true continuation, Director extracts the final frame from the selected predecessor video and stores it as a hidden project image. The successor shot uses that frame as its visual starting point and records the exact predecessor generation that supplied it.

Some video models cannot combine first-frame conditioning with a separate stack of visual references. In that case the predecessor frame becomes authoritative, while the prompt restates the named identity and setting anchors. The system should make that tradeoff explicitly instead of sending conflicting inputs.

Director is agent-native, but it is not agent-only. A creator can ask the Director Agent to build or revise the storyboard, then directly rename a reference, add a scene, move a shot, or replace an image. Both paths call the same project services.

That shared state matters. If chat and direct manipulation become parallel workflows, creators cannot tell which plan is current. In Director, a successful agent change refreshes the visible panels, and a direct edit becomes part of the next conversation context.

Planning can change freely, but image and video generation can spend real credits. Director keeps those concerns separate. The agent may prepare the complete plan, calculate the missing work, and explain the order, but a person approves the paid runs.

The approved plan is durable. References can run first, independent shots can follow, and continuation shots wait for the selected predecessor result. Reloading the page should not turn an approved sequence into a forgotten chat instruction.

In a bounded two-shot staging pilot, the most important evidence was not that both clips completed. It was that the second run could be traced back to the first selected video, its extracted final frame, the same named world references, and the intended next end state.

The pilot also reinforced a product lesson: continuity is not something to hide behind one Generate button. Creators need to see the world, the scenes, the shots, the selected results, and the approval boundary, while the agent handles the coordination between them.