Generating Video Scripts with Gemini AI

Both legacy scenes schema and modern slides+highlights schema are valid, depending on your pipeline.

Approach A: Legacy Scene-based Schema

Good for standalone script generation or older renderers.

{
  title: string;
  description: string;
  scenes: Array<{
    sceneNumber: number;
    headline?: string;
    narration: string;
    visualDescription: string;
    duration: number;
  }>;
}

Approach B: Pipeline Schema (Slides + Highlights)

Good for modern multi-stage pipelines.

{
  videoFormat: 'long' | 'single_short' | 'multi_short';
  title: string;
  description: string;
  thumbnailDescription: string;
  highlights: Array<{ title: string; description: string; reason: string; slides: number[] }>;
  slides: Array<{
    slideIndex: number;
    headline: string;
    imageDescription: string; // English
    audioNarration: string;   // Japanese
    estimatedDuration: number;
    directorNotes?: string;
    audioProfile?: 'urgent' | 'calm' | 'excited' | 'serious' | 'casual' | 'dramatic';
  }>;
}

Prompting Rules (Works for Both)

Force JSON-only output.
Keep language boundaries explicit.
Ground facts in source material.
Include narration and visual guidance per unit (scene/slide).

Normalization and Ownership Rules

In stage-based pipelines:

selection stage owns final format choice
script stage can normalize model output to selected format
script save should avoid overwriting selection-owned fields (video_format)

Cost Tracking