How do you break a long screenplay into scenes for AI video production without losing story continuity?

Break the screenplay at its own seams — acts, then scenes, then shot clusters of 3–5 frames sharing a location, character set, and lighting state — and lock a single production bible (script, characters, world, tone) into a creative producer agent that every downstream agent reads from. That shared context is what keeps continuity intact across chunks.

Start by loading the FULL script into a creative producer agent before you chunk anything. The invideo agent is an agentic video tool with all current video and image models routed behind it, so this producer agent becomes the central vision-holder — script, shot breakdown, character details, tone — that every other agent (storyboard, DOP, costume, production design) reads from. Continuity problems on long scripts almost always trace back to chunks being generated against partial context; front-loading the whole script solves it once.

Chunk at story seams, not time lengths. Split the screenplay into acts first, then scenes inside each act, then shot clusters of 3–5 frames that share a location, character set, and lighting state. Never cut chunks at arbitrary durations — a chunk should be a unit the agent can hold one consistent visual logic across. One documented production worked 25% at a time, fully completing storyboarding, generation, and editing for one act before starting the next: "I'm not overworking the AI where it kind of loses context down the line. I like to lock in on something and then move forward."

Lock the world and the cast before you generate a single scene. Day 1 of any long-script production is pre-production inside the invideo agent: generate 4 options per character sheet and per environment reference, pick one, and lock them. Across documented productions, this step ran 11 image generations for 4 characters and 1 prop on one project, and used four-option grids on another. After locking, those approved panels REPLACE the original references — every subsequent chunk pulls from the locked sheets, which is what holds character identity and world look across acts. No LoRA needed; one 70-second short held 2 characters consistent across every scene this way.

Build a short production bible the agent reads as system context. A compressed document — characters (with sheets attached), locations, tonal palette, lens grammar, story arc position per act — is what the producer agent hands to every sub-agent. One production used a 25-page treatment encoding 14 sections (camera, lighting, palette, composition, mood, negative prompts) and a 9-element prompt assembly order held across every frame. You don't need 25 pages; you need enough that the agent can ask, for any new chunk, "where in the arc are we, who's in the room, what's the light doing?" and answer it without you re-explaining.

Run agents in parallel, one chunk at a time. Assign a storyboard agent to visualize the chunk first, then a DOP agent (one per scene — different scenes need different eyes; documented productions ran 2 DOPs on a single complex scene and 6–8 specialist agents simultaneously across a project) to design the shots, then go to generation. Because every agent inherits the producer agent's context, characters and world stay locked even when you're parallelizing across cities and people — one production ran 3 people across 3 projects simultaneously with no continuity break.

Carry visual continuity ACROSS chunk boundaries explicitly. At the end of each chunk, log everything generated back to the producer agent — approved frames, any manual overrides, any character-sheet corrections — so its memory of the film stays current. When you start the next chunk, a minimal continuation prompt like "Everything should match" is enough for the agent to maintain character, lighting, lens grammar, and spatial logic, because the locked sheets and bible are still loaded. If a continuity error appears, fix it at the SOURCE in the character sheet — the agent identifies the exact panel, corrects it, stores the update, and downstream chunks inherit the fix automatically.

Close each act with a maker-checker pass. Send the act's rough cut back to the invideo agent with an open "what's working, what's not" prompt against the loaded bible. On one production this caught the entity-reveal running at the wrong emotional stage register — the kind of structural arc error that's invisible chunk-by-chunk but obvious against the full script. Then move to the next act.

Across documented productions running this pattern, films of 70 seconds to 7 minutes have been produced in 2–5 days at $315–$750 per finished minute, with character and world continuity holding across 21+ scenes in a single project.

Watch some of these to see what works for you:

To really set up the context for the agent, I normally start off with the creative producer agent. That's where I'll give the script, or the shot breakdown, along with the characters. That's the main agent that sort of holds the understanding and the vision of the entire film.

— Hridaye, invideo's creative director