Seedance Storyboard

A timeline storyboard flow for Seedance. Largely covered by Seedance 2.0 Prompting — read this page before installing it; its value is six effects-replication and commercial examples the other skill lacks.

Install Seedance 2.0 Prompting instead — unless you are doing effects replication or commercial work.

We scored these two as complementary. Running them against each other proved that wrong, and this page exists to say so rather than to quietly recommend both.

Where they overlap

The platform specifications are identical, item for item: image and video and audio limits, file counts, generated duration, pixel range. Not similar — the same numbers. And on the thing that matters most for this model, camera movement, its guidance is roughly an order of magnitude thinner than the other skill’s four-axis system.

Installing both is actively worse than installing one. An agent asked to write a Seedance prompt would be holding two sets of instructions that disagree.

What it does have

Six worked examples in a domain the other one does not cover: Hitchcock zoom replication, character-consistent narrative, video extension, costume-change effects, a finished commercial, and a single-take multi-scene transition.

The other skill’s 31 examples are almost all live-action social video. These six are effects replication and brand work — which is closer to what we actually make. That is the one real difference, and it is a difference in corpus, not in method.

Our own conclusion: take the six examples and fold them into the other skill’s reference set, rather than installing this one whole.

What we ran

A fifteen-second, five-shot timeline storyboard through its five-step flow. The output passed the same checker we used on the other skill — 545 characters, four camera verbs, naming conventions correct, no-real-faces declarations present. It works. It is just covered.

The fifteen-second, five-shot storyboard the skill produced
The complete output: five steps, a timeline table, the prompt in its template structure, and the self-check it runs at the end.

What we found running it

Three tasks, each run with the skill installed and again with skills disabled, same task text. Both arms produced a workable storyboard every time. The split is over what comes out alongside it.

The skill arm shipped exactly two files each run: a Chinese storyboard document and the finished Seedance prompt lifted out into its own file — seedance_prompt.txt (1,469 B and 1,141 B) and 茶饮广告_Seedance提示词.md (3,990 B) — so the thing you paste is not buried inside the markdown. Its storyboards ran longer as well (6,405 B / 5,390 B / 7,548 B against 5,019 B / 3,781 B / 4,764 B), and all three carried step-by-step instructions for the Jimeng platform specifically. The control arm spent its extra effort on machine-readable formats instead: shotlist.csv, a storyboard.json (4,243 B) with a per-shot English prompt field, and in task 3 a separate negative_prompt.txt (1,368 B).

Task 3 is where the difference becomes substantive. The brief asked for a single unbroken take — 不要那种一段一段切开的镜头感. The skill arm designed four invisible transition points, naming the technique at each (motion match, shape match, direction continuation, liquid topology continuity) and specified that the whole 12 seconds be produced from one prompt in one pass. The control arm split the shot into five scenes and relied on physical transitions to hide the joins — a good answer to a different question, since a segmented generation is exactly what the brief ruled out.

One concrete reliability difference: in task 1 the control arm's answer lists three deliverables including storyboard.md, but only shot_prompts.md (5,019 B) and shotlist.csv (1,014 B) exist on disk. The storyboard it describes was never written. The skill arm's file lists matched what it shipped in all three runs.

Cost was near-identical: $0.25 / $0.26 / $0.30 with the skill against $0.23 / $0.21 / $0.27 without.

What we did not test: the videos. We never fed a single one of these prompts to Seedance, so the central claim — that the one-pass prompt in task 3 yields a genuinely unbroken shot — is unverified. The Jimeng platform steps are likewise untested. Three tasks, all Chinese, all between 10 and 15 seconds.

Installing

If you still want it, install from the source repository. MIT. Note that the repository URL printed in its own documentation is dead — elementsix/claude-seedance-plugin does not exist; the skill lives in the repository linked above.

Related Skills

Explore More →