Seedance2-Storyboard-Generator: A Claude Code Skill That Turns Novels Into Multi-Episode Video Scripts

The blank page problem in AI short drama, not the rendering problem

Seedance2-Storyboard-Generator does not generate video. Nothing in the repository talks to a model API. What it ships is a Claude Code skill plus a set of Markdown conventions that convert a novel or story into three artifacts: a four-act script, a numbered asset list, and per-episode storyboard prompts written in Seedance 2.0's timeline format. The README frames the motivation bluntly: trial and error is getting more expensive, and prompt quality matters more than it used to. That is the actual problem being solved. Once a generation costs real money and real minutes, a director who improvises each prompt pays for every mistake. The skill front-loads the decisions (who the characters are, which scene, which prop, which camera move in which three-second slot) into text you can review before spending anything. The intended user is someone adapting prose into a multi-episode vertical series, working in Chinese, already inside Claude Code. The repository's own examples are all Chinese-language adaptations: a Lin Chong arc from Water Margin, a Nie Feng arc from Fung Wan, and Maupassant's The Necklace. If your source material is an English novel and your target platform is not Seedance 2.0, the conventions here will not map cleanly.

Four acts, then C/S/P numbering, then a 15-second timeline

The pipeline in the README is a straight line: idea, script, asset descriptions, image generation, storyboard script, per-episode video. The mechanism that makes it more than a checklist is the numbering scheme. Characters take the prefix C, scenes take S, props take P, each with a two-digit range from 01 to 99. So C01 might be a character's front full-body view, S01 a snow scene at a specific location, P01 a spear. Those identifiers then reappear inside the storyboard prompts as reference slots: a table maps 图片1 to C01 and 图片2 to S01, and the prompt body refers to them with @图片1 and @图片2. That indirection is the whole design. It means a character can be revised once and every downstream storyboard that cites C01 inherits the change, at least in the text layer. The storyboard itself is sliced into three-second beats across a fifteen-second episode: 0-3 seconds for an establishing aerial shot, 3-6 for pushing in on the subject, and so on through a closing frame. Below the visual beats sit two labelled blocks, one for sound (score style, effects, dialogue) and one for references. Each episode file also records a tail-frame description, the final frame of the episode, which the next episode uses as its continuation point. Style consistency is handled by a mandatory prefix on every asset prompt; the README gives 'Chinese ink wash painting style mixed with anime cel-shading' as the example, with character differentiation pushed onto colour schemes and visual markers rather than prompt wording.

Invoking the skill and the file naming it expects

The README's usage section is one line: typing 'seedance' followed by a story premise, for example 'seedance 风云中聂风小时候的故事', triggers the skill. The skill handles story analysis, four-act structure, asset planning, and Seedance 2.0 format prompts. The skill definition lives at .claude/skills/seedance-storyboard-generator/SKILL.md, with two reference documents beside it: references/seedance-manual.md and references/故事转视频脚本-转换工具.md. The repository root holds a folder named 写剧本和分镜 containing the .claude directory, a docs folder, three example projects, and an asset folder. File naming is fixed by convention rather than by tooling: scripts are [title]_剧本.md, asset lists are [title]_素材清单.md, and episode storyboards are [title]_E[number]_分镜.md. That convention matters more than it looks. Because the skill reads and writes by pattern, a renamed file breaks the chain silently. There is no build step, no package manifest, and no installer described in the README. You clone the repository, place it where Claude Code can see the skill, and the rest is Markdown. The documentation also points to an online version hosted at skillhub.cn, described as free to use directly, which is an alternative to local setup but puts your source material on someone else's service.

The platform caps that shape every prompt you write

The limitations section is the most useful part of the README because it constrains the design rather than describing it. Seedance 2.0 accepts at most nine images per generation and at most three reference videos totalling fifteen seconds. Sensitive words can fail a generation, and the platform does not tell you which word triggered the block; the README's own troubleshooting advice is a bisection method, splitting the prompt and testing segments to narrow the range, plus maintaining a personal blocklist. Video editing is described as limited, meaning most corrections require a full regeneration rather than a patch. Instruction-following degrades on complex prompts, with 300-plus characters called out as the point where consistency suffers. That last cap is in direct tension with the rest of the workflow: the timeline format is verbose by construction, five time slots plus sound and reference blocks. A prompt that specifies camera movement, character reference, scene reference, mood, and dialogue can pass 300 characters quickly. The README does not resolve this tension. It states both the format and the ceiling and leaves the trimming to you. Anyone adopting this should treat the 300-character figure as the real budget line, not the fifteen-second episode length.

Video extension as the continuity mechanism, and where it fails

Episode chaining works through Seedance's video extension rather than through re-generation. From episode two onward, the previous episode's video is uploaded as @视频1 and the prompt opens with an instruction to extend it by fifteen seconds. The README diagrams the resulting chain: a newly generated title sequence, a newly generated E1, then E2 as an extension of E1, E3 as an extension of E2. The appeal is continuity of lighting, grain, and character appearance without re-specifying them. The failure mode is stated in the FAQ: extended transitions can look unnatural, and the suggested fix is to write the transition into the prompt explicitly, for example a fade to black with the sky shifting from night to dawn. That is a text-level workaround for what is fundamentally a model behaviour, and it will not always hold. There is a second, structural cost the README does not discuss. Because each episode extends the last, an error early in the chain propagates forward. A character who drifts in E2 is baked into E3 and E4. The four-act script and the tail-frame notes are the only defence, and they are both text artifacts that a model may or may not honour across a fifteen-second extension.

What to compare it against, and why the comparison is not close

The obvious alternative is writing Seedance prompts by hand from a screenplay, which is what most people do now. The difference is not quality, it is bookkeeping. A hand-written workflow keeps character descriptions in your head or scattered across documents; the C/S/P scheme keeps them in one numbered namespace that every storyboard references. The second alternative is a general-purpose storyboard tool such as a shot-list template in a screenwriting application. Those tools are built around human crews and human shot language: coverage, eyelines, lens notes. They have no concept of a reference-image slot or a three-second generation window, so the output needs a translation pass before it is useful to Seedance. The third alternative, and the one worth naming plainly, is that Seedance 2.0's own platform interface plus a well-kept personal prompt library may be enough. If you produce one video a month, the skill's overhead (learning the file naming, maintaining the asset numbering, keeping the style prefix consistent) likely exceeds the benefit. The repository's advantage compounds with volume and with team size, because the numbering scheme is what lets a second person pick up episode four without re-reading episode one.

Licence, maintenance, and what the repository does not tell you

The licence is the weakest point in the material. The README's final line states that the content is for study and reference only. That is not a licence identifier, and the repository metadata retrieved here lists no licence at all. For a project whose output is intended for commercial short-drama production, that ambiguity matters: you cannot tell from the README whether generated scripts and asset lists carry any restriction, or whether the bundled example projects (Lin Chong, Nie Feng, The Necklace) can be reused. The examples themselves are adaptations of public-domain and copyrighted sources, and the README does not address that either. Maintenance cost is low in the ordinary sense, because there is no code to compile and no dependency to update. The real maintenance burden is external: Seedance 2.0's prompt syntax, image limits, and sensitive-word behaviour can change without notice, and the skill's reference manual is a static Markdown file that someone has to keep in sync. The repository shows no releases, so there is no version tag to pin against. Treat the docs folder as the thing to watch, and re-read references/seedance-manual.md after any platform update rather than assuming the skill still matches the current model.

Editorial conclusion

Adopt it if you already write in Claude Code and produce 9:16 vertical short drama in Chinese, because the C/S/P numbering and the timeline prompt template remove the blank-page problem. Skip it if you need a renderer, an English-language pipeline, or per-shot control, since the repository only emits prompts and its own README caps complex instruction-following at roughly 300 characters. Before committing, open .claude/skills/seedance-storyboard-generator/SKILL.md and the 林冲项目 sample files to check whether the output granularity matches your editing process, and confirm the licence with the author, because the README states only that the content is for study and reference.

Official sources

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