Seedance2-Storyboard-Generator: a Claude Code skill for turning novels into multi-episode video scripts

The problem: prompt craft replaced production budget, and nobody wrote the script step down

Generating a single good clip is a solved reflex. Generating twenty of them that look like the same production is not. The gap is not the video model, it is everything upstream: who the characters are, what they wear, which prop sits in which shot, and how episode four continues from the last frame of episode three. Seedance2-Storyboard-Generator targets that gap. The README states the motivation plainly: trial and error is getting more expensive, and prompt quality matters more than it ever did.

The audience is narrow and specific. You are making short-form serialized drama, likely vertical, likely from an existing text (a classic novel chapter, a short story, a folk tale). The repository's own sample projects are exactly that: 林冲 from Water Margin, 聂风 from Fung Wan, 项链 from Maupassant, plus 司马光, 岳阳楼记, 崖山海战 and 鲁智深 projects. It is a Chinese-language workflow for Chinese-language source material. If your source is English and your target platform is not Seedance, the naming conventions and the prompt manual will fight you.

What actually runs: a Claude Code skill, not a program

There is no server, no CLI binary, no package to install. The mechanism is a Claude Code skill directory at .claude/skills/seedance-storyboard-generator/, containing SKILL.md as the definition and workflow, a references/seedance-manual.md with the full Seedance 2.0 manual, and a story-to-script conversion template. Claude Code reads that directory and gains the ability to run the workflow when invoked.

The pipeline is documented as a linear chain: 构思主题 → 写剧本 → 生成素材描述 → 生图 → 写分镜脚本 → 逐集生成视频. Concretely, the skill adapts source material into a four-act structure (起承转合), then produces numbered asset prompts, then produces per-episode storyboard files. Image generation is deliberately outside the skill: the README names GPT-Image-2, Seedream and Nana Banana Pro as options and says other image AIs are acceptable. Video generation happens on the Seedance 2.0 platform, not here.

The naming scheme is the part that makes the output machine-checkable. Scripts are `[标题]_剧本.md`, asset lists are `[标题]_素材清单.md`, and episodes are `[标题]_E[集数]_分镜.md`. Assets are prefixed C for characters (C01-C99), S for scenes (S01-S99), P for props (P01-P99). Each storyboard file carries three blocks: an asset slot table mapping image slots to C/S/P codes, a Seedance prompt in timeline format, and a tail-frame description of the episode's last frame. That tail frame is the continuity contract between episodes.

Installing and running the skill inside Claude Code

The README gives no package manager step because there is nothing to install in the usual sense. The skill lives in the repository, so you clone or download it and open the project directory in Claude Code. The layout in the README shows the workflow folder 写剧本和分镜/ containing the .claude directory, docs, and the sample projects.

Once the directory is present, the README's usage example is a single invocation line. The skill then handles story analysis, four-act structure, asset planning and Seedance-format prompts.

❯ seedance   风云中聂风小时候的故事

What you should end up with is a set of markdown files following the naming convention: a script, an asset list, and one storyboard file per episode. The README points to 林冲项目/使用指南.md as a worked example if the output shape is unclear.

The storyboard prompt itself has a fixed shape. The README shows a 15-second vertical 9:16 prompt with a style line, five three-second beats, and bracketed sound and reference blocks:

水墨武侠风格，9:16竖屏，灰白墨黑色调

0-3秒画面：高空俯拍，场景建立
3-6秒画面：镜头推进，主体引入
6-9秒画面：情节发展，冲突展现
9-12秒画面：高潮时刻，情感爆发
12-15秒画面：结尾落版，氛围延续

【声音】配乐风格 + 音效 + 对白
【参考】@图片1 角色，@图片2 场景...

Episodes two onward are not generated fresh. The README instructs you to upload the previous episode as @视频1 and open the prompt with `将@视频1延长15s`, producing the chain 片头(新生成) → E1(新生成) → E2(延长E1) → E3(延长E2).

The hard limits that decide whether your story fits

The README's 重要限制 section is the most valuable page in the repository, because it is the part that will kill your project if you ignore it. Seedance 2.0 accepts at most 9 images per generation and at most 3 videos totalling 15 seconds as reference. With a C/S/P numbering scheme that runs to 99 per category, the cap is not theoretical: a scene with three characters, a location and two props is already six of your nine slots, before you add angle variants.

The second limit is instruction adherence. The README states that complex prompts of 300+ characters may be followed inconsistently. The timeline format above is roughly that length in Chinese before you add sound and reference blocks, so the format sits right at the edge of what the model tracks reliably.

The third is sensitive-word blocking. The README is explicit that the platform does not tell you which word triggered the block, and recommends a bisection approach: split the prompt into segments, test them, narrow the range. It also suggests building a personal blocklist, naming 江湖人士 as an example. That is a manual, iterative debugging loop with no error message, and it is the single biggest source of wasted generation attempts in this workflow.

Finally, video editing is limited. The README says most modifications require regeneration rather than an edit. If your process assumes you can fix a bad shot in post, this workflow does not support that assumption.

Style consistency is a prefix convention, not a feature

The README's answer to visual consistency is a string convention: every asset prompt must begin with the same style prefix, with `Chinese ink wash painting style mixed with anime cel-shading` given as the example. Character recognisability is handled by different colour schemes and visual markers rather than by any model-level identity mechanism.

That is a real design choice with a real cost. A prefix convention holds up when the same model generates every asset. The README permits mixing image generators (GPT-Image-2, Seedream, Nana Banana Pro, or others), and the moment two different generators render the same prefix, the prefix stops guaranteeing a shared look. Nothing in the repository enforces or verifies the prefix; it is a discipline the writer has to maintain across every asset prompt by hand.

The camera-movement vocabulary is a similar artefact. The README supplies a keyword table mapping effects to Chinese terms: 推镜头 for push-in, 拉镜头 for pull-out, 环绕镜头 for orbit, 希区柯克变焦 for the dolly-zoom effect, 手持晃动 for handheld shake. It is a useful crib sheet precisely because these are the tokens the platform responds to, but it is documentation, not a validated grammar.

Where it is the wrong tool, and what to use instead

This is the wrong tool if you want automation. There is no API surface, no batch runner, no headless mode. Every episode passes through a Claude Code conversation and a manual upload to the Seedance platform. For a five-episode 林冲-style series that is fine. For a hundred-episode production it is a bottleneck with no obvious escape hatch in the repository.

It is also the wrong tool if you need English output. The skill files, the prompt manual, the templates and every sample project are in Chinese, and the sensitive-word problem is specific to a Chinese-language platform filter. Translating the workflow means rewriting the reference documents, not just the README.

The honest alternative is not another repository, it is doing the work by hand with a general-purpose chat model. A four-act script and a numbered asset list are things any capable model can produce from a prompt; what this project adds is the fixed file naming, the C/S/P numbering, the 15-second five-beat timeline template, and the tail-frame continuity convention. If you already have your own conventions and a style guide that works, the skill's structure is overhead rather than help. If you do not have those conventions, adopting a documented one is faster than inventing one, which is the actual case for this repository.

Maintenance, licence and what the repository does not say

The last push was on 2026-09-06, and the repository is not archived. The README states the skill is live and free to use at a skillhub.cn URL, and that the project was first posted on the linux.do community. There are no retrieved releases, so there is no version history to reason about and no changelog to check before upgrading.

The licence is the weak point. The README's 许可 section says only that the content is for learning and reference. No licence identifier is given, and the repository's licence field is unknown. That is not a permissive open source grant, and it is not a clear prohibition either. If you plan to build a commercial series on top of these templates, that ambiguity is worth resolving with the author before you invest in a multi-episode run, and it is a question for a lawyer rather than for this article.

Upgrade cost is low by construction: the artefact is markdown, so pulling a newer SKILL.md or seedance-manual.md does not break anything you have already generated. The real cost sits on the other side of the boundary, in Seedance 2.0 itself. If the platform changes its image cap, its reference-video rules or its filter behaviour, the README's limits section goes stale and nothing in the repository will tell you.

Editorial conclusion

Adopt it if you already run Claude Code, write in Chinese, and are producing short multi-episode drama where a 15-second-per-episode rhythm fits. Skip it if you need an API, a renderer, or English-language output, because none of those are in the repository. Before committing, open .claude/skills/seedance-storyboard-generator/SKILL.md and read the 9-image and 3-video limits against your own shot list; those caps decide whether your story fits the format at all.

Frequently asked questions

Official sources

1. Issues
2. liangdabiao/Seedance2-Storyboard-Generator on GitHub
3. Project website
4. README