← All work
Creative AI 2026

Story Engine

One subject in, a consistent narrated video out.

The problem

Generating multi-scene illustrated video from a topic normally suffers character and style drift across frames — the same person looks like three different people.

How it works

A CLI pipeline where Claude first generates a locked Story Bible — a narrative arc, named characters with head-to-toe locked descriptions, locked locations, and an opinionated style token — which is injected verbatim into every downstream image prompt to enforce consistency. Scenes are generated as constrained enums over the character and location IDs; images via Replicate, narration via ElevenLabs, final cut in Remotion.

Key features

  • Locked Story Bible for cross-scene character and style consistency
  • Zod-schema-constrained structured generation
  • Replicate image generation + ElevenLabs narration
  • Remotion (React) final video render

Architecture

TypeScript modules pipe bible → scenes → image prompts → images → audio, with JSON metadata and assets feeding a Remotion composition for the render.

Stack

TypeScript Claude (Anthropic SDK) Zod Replicate ElevenLabs Remotion