What it does

This is a workflow template and Claude Code skill for converting written stories into serialized AI video. You feed it a story; it outputs a four-act screenplay, numbered asset lists (characters, scenes, props), and timed shot-by-shot prompts formatted for Seedance 2.0. The repo includes three complete example projects — a Water Margin adaptation, a Feng Yun episode, and Maupassant’s “The Necklace” — with full scripts, asset manifests, and per-episode storyboards.

The interesting bit

The real craft is in the constraints. The author treats Seedance 2.0’s limits — 9 reference images max, 3 video references, 15-second clips, mystery sensitive-word filters — as creative parameters rather than obstacles. The workflow explicitly tracks “last-frame descriptions” to chain episodes via video extension, turning a fragmentary generation process into something resembling continuity editing.

Key highlights

• Structured prompt format with 3-second time-slice granularity (0–3s establishing shot, 3–6s subject introduction, etc.)
• Asset numbering convention: C01–C99 characters, S01–S99 scenes, P01–P99 props
• Video-extension chaining: E2 extends E1, E3 extends E2, maintaining (in theory) visual continuity
• Includes a 20-episode ink-wash anime project as reference implementation
• Also published as a free Coze skill for non-technical users

Caveats

• Heavy reliance on Seedance 2.0’s proprietary platform; workflow breaks if the tool changes
• “Seamless transition” is aspirational — the FAQ explicitly addresses fixing jarring extensions with prompt hacks like “画面渐暗过渡”
• Complex prompts over 300 characters may be ignored; sensitive-word filtering is opaque and requires binary-search debugging

Verdict

Worth studying if you’re building structured AI video pipelines or trying to squeeze narrative coherence out of short-form generators. Skip if you want a general-purpose video tool — this is tightly coupled to Seedance 2.0 and Chinese-language short-drama formats.