The agent first reconstructs what it can verify from the reference (hook, goals, escalation, ending, with timestamps) and notes the narrative technique worth borrowing, then writes a new story rather than a reskin. It lays out characters, costumes and locations, builds a timed shot table, and groups shots into generation segments under a 15 or 30 second cap chosen by the user, splitting early at scene or costume changes and stating each segment's start and end state. Every segment gets a self-contained video prompt, and every character, outfit and scene gets an image prompt. It only plans, with no model calls or API keys, and the README documents installation for Codex only, with other agents unverified; the instructions are in Chinese.