seedance 2.0 repo prompt generator

generate seedance 2.0 prompts by analyzing the current repository (README, routes/screens, ui strings, assets) and turning that into a 30s montage pack.

when to use

use this skill when the user wants:

• motion graphics videos for their app / saas / program
• a repo-driven promo script + prompts (no manual prompt writing)
• a 30s cut built from "best of" generated clips

core idea: seedance is the engine

seedance is best treated as the engine, not the vehicle. ship results by:

1. generating more than you need
2. selecting the best 1-2 seconds from multiple runs
3. cutting down to a tight final

the op montage method (default)

• write a 30s script
• split into 5 segments x 6s (final)
• for each segment, generate a 15s mini-montage prompt (overshoot)
• keep the best of the best clips
• cut back down to 30s

keeper rule:

• put the hero moment in 00:00-00:06 of each 15s generation
• use 00:06-00:15 as coverage (alternate angles / salvageable micro-moments)

repo analysis workflow (do this every time)

1. detect project type (web app, mobile app, cli, library)

• check for package.json, pyproject.toml, Cargo.toml, go.mod, etc.

1. extract product narrative

• read README.md (and docs/, CHANGELOG*)
• pull 3-7 crisp product claims: what it does, for whom, why it matters

1. identify showable surfaces

• web apps: pages/routes + nav labels + key components
• clis: key commands/flags + representative output strings
• libraries: public api + short usage snippet

1. locate existing visual assets

• prefer repo images: assets/, public/, docs/, screenshots/
• if none: output a screenshot checklist the user can capture in <10 minutes

1. choose a motion style that matches the repo

• if brand cues exist (palette/typography), align
• otherwise default: clean professional product film

seedance constraints to respect

• inputs: max 9 images (@Image1..@Image9), max 3 videos, max 3 audio
• do not rely on tiny text / pixel-perfect ui
• avoid real-person faces and celebrity likenesses

filter-aware prompting (37% fail rate)

the content filter is an LLM that reads context, not keywords:

• lead with cinematic language: "cinematic wide shot, 35mm film grain, 16:9" → creative context = higher pass rate
• describe characters by role, never age: "rider" not "young boy", "figure" not "child"
• build context around sensitive elements: wrap actions in narrative/film language
• faces in image inputs = auto-reject: crop to back of head/shoulders or use wide shots
• every sentence earns its place: either describe what camera sees OR anchor as cinematic

output contract

return exactly this structure:

## seedance montage pack: <project name> (30s master)

### repo-derived notes
- product: ...
- audience: ...
- key features/screens to show: ...
- brand/style cues detected: ...

### asset plan (upload order)
@Image1: ...
@Image2: ...
@Image3: ...
@Video1 (optional): ...
@Audio1 (optional): ...

### 30s script (5 x 6s)
- segment 1 (0-6s): ...
- segment 2 (6-12s): ...
- segment 3 (12-18s): ...
- segment 4 (18-24s): ...
- segment 5 (24-30s): ...

### segment prompts (15s mini montages)

#### segment 1 prompt (recommended)
<paste-ready seedance prompt>

#### segment 1 prompt (alternate)
<paste-ready seedance prompt>

... repeat for segments 2-5 ...

### assembly notes
- generation plan: ...
- selection rubric: ...
- editing plan: ...

### variants
- 9:16 notes: ...
- 16:9 notes: ...