longform-cinematic-video

Direct one film, not a collection of clips

Own the production-level decisions that cross model generation and editing. Read video-storytelling, video-cinematography, minimax-h3-video, video-provider-capabilities, and video-editing for their specialist rules. Keep those responsibilities coordinated by one sequence plan and one editorial grammar.

Define the cinematic contract

The user chooses the visual style: photoreal, hand-drawn, 2D/3D animation, motion graphics, documentary, slideshow, or mixed media. Cinematic means coherent storytelling and intentional picture/sound direction, not realism. A photoreal host may have illustrated historical cutaways. Do not refuse or convert those inserts to photoreal. Realistic characters are not required.

Before scripting, lock the film's dramatic question, audience, runtime, format, emotional arc, point of view, visual language, camera grammar, color journey, sound world, and continuity priorities. State what must remain exact across the film: character identity, wardrobe, props, geography, time of day, screen direction, eyelines, lighting motivation, and recurring audio motifs.

Choose a small visual vocabulary and repeat it intentionally. A cinematic film does not need a new style, lens, camera move, or transition for every beat. Use the user-approved medium for each sequence; a film may combine generated footage, supplied media, and deterministic animation. H3 reference and continuation rules apply to generated-video shots, not every visual sequence.

Plan sequences before shots

Organize the film as chapters, sequences, scenes, and only then shots. Give each sequence an entrance state, dramatic turn, exit state, location, characters, time, lighting, sound, and continuity bridge into the next sequence. A script beat is not automatically a new generated clip.

Create longform-sequence-plan.json before paid video generation. Include:

• chapter, sequence, scene, and shot identifiers;
• measured narration or dialogue duration covered by each sequence;
• H3 route chosen by actual controls: text-only, first-frame or minimal references;
• references and their exact semantic roles;
• incoming and outgoing character, prop, geography, motion, camera, lighting, and audio state;
• generation topology: one H3 take, Reference-to-Video continuation, a motivated H3 camera-angle change, or intentional hard cut;
• planned cut point, transition grammar, handles, and expected seam risk;
• the reason every independent generation is unavoidable.