AIVideo/.codex/skills/novel-to-seedance-video/SKILL.md

---
name: novel-to-seedance-video
description: Convert a user-specified Chinese novel project folder or file into a fully automated multi-episode AI video workflow for Seedance 2.0, MJ and nano-banana-2 image assets, and Jimeng/Dreamina, including story analysis, episode segmentation, per-episode storyboard, rich Chinese character/scene/prop asset prompts, official-site @ uploaded image prompts, API payloads, platform safety review, and per-step output files. Use when the user asks to turn a specific fiction, web novel, chapter folder, script folder, or story text file into AI video shots or a reusable novel-to-video production process.
---

# Novel To Seedance Video

## Use the project workflow

Follow `workflows/novel_to_seedance/01_WORKFLOW.md`.

Read references only as needed:

- `workflows/novel_to_seedance/references/input_project_selection_rules.md` for user-specified date/project folder input handling and output mapping. Do not auto-process all input projects.
- `workflows/novel_to_seedance/references/image_model_provider_rules.md` for outputting both MJ and nano-banana-2 asset prompts and optional model recommendations.
- `workflows/novel_to_seedance/references/seedance2_prompt_rules.md` for Seedance prompt structure.
- `workflows/novel_to_seedance/references/midjourney_asset_rules.md` for MJ character, expression, scene, and prop prompts.
- `workflows/novel_to_seedance/references/rich_asset_prompt_rules.md` for high-density Chinese-only character sheets, multi-view scene boards, and multi-angle prop sheets.
- `workflows/novel_to_seedance/references/performance_naturalism_rules.md` for reducing AI-looking expressions, gestures, overacting, and stiff video performance.
- `workflows/novel_to_seedance/references/legacy_storyboard_notes.md` when the user mentions the old first-episode storyboard.

## Output rule

For each user-specified novel project, create a corresponding output folder under:

`outputs/novel_video_runs/date-batch/project-name/`

Output global files first:

1. `00_source_intake.md`
2. `01_series_analysis.md`
3. `02_segmentation_decision.md`
4. `03_series_bible.md`
5. `04_platform_safety_rules.md`

Then create one folder per episode under `episodes/EP##_episode-name/`, and for every episode output:

1. `01_episode_analysis.md`
2. `02_adaptation_plan.md`
3. `03_storyboard.md`
4. `04_character_scene_delta.md`
5. `05_image_asset_prompts.md`
6. `06_jimeng_asset_prompts.md`
7. `07_seedance_prompts.md`
8. `08_transition_snapshots.md`
9. `09_prompt_safety_review.md`
10. `10_qc_checklist.md`

After all episodes are complete, output `99_series_qc_summary.md`.

## Core method

1. Analyze plot before writing shots.
2. Only process the user-specified input folder or file. If multiple input projects exist and the user did not specify one, ask which project to process.
3. Select high-value cinematic beats: conflict, reveal, choice, reversal, danger, rescue, humiliation, counterattack.
4. Convert internal narration into visible action, props, expression, sound, and spatial blocking.
5. Keep each Seedance segment within 4-15 seconds.
6. Split long actions into multiple segments; merge tiny beats into one multi-shot segment.
7. Design characters and scenes before final video prompts.
8. Output both MJ and nano-banana-2 image asset prompt versions unless project config explicitly asks for only one.
9. Execute all recommended episodes by default. Do not stop after episode 1 unless the user explicitly requests a single episode.
10. Use stable `@角色`, `@场景`, and `@道具` names in every Seedance prompt.
11. Write transition snapshots for every segment and between episodes.
12. Run prompt safety review for MJ, nano-banana-2, Jimeng/Dreamina, and Seedance before finalizing prompts.
13. End with per-episode and series quality checklists.

## Platform safety review

For every MJ/nano-banana-2/Jimeng/Seedance prompt:

1. Classify risk as `OK`, `LOW`, `MEDIUM`, `HIGH`, or `BLOCKED`.
2. Rewrite `MEDIUM` and above before final output.
3. Avoid explicit adult/sexual content, sexualized minors, gore, graphic injury, self-harm, dangerous instructions, illegal activity details, hate, harassment, misinformation, real-person deception, privacy/IP violations, readable official documents, and sensitive political content.
4. Prefer safer ancient-drama wording: `暗害` over explicit killing, `危险药瓶` over specific illicit drugs, `倒下` over graphic death, `逼退到墙边` over graphic assault, `明黄卷轴` over readable decree text.
5. State that safety review reduces risk but cannot guarantee platform approval.

## Style standard

Prefer cinematic continuity over isolated pretty shots. Keep prompts executable: concrete camera, action, expression, light, sound, and tail frame. Avoid abstract emotions unless paired with visible behavior.

For performance, prefer natural actor behavior: small pauses, eye-line shifts, breath changes, incomplete gestures, and restrained micro-actions. Avoid exaggerated eyes, cold smiles, stage-like posing, and abstract emotion labels.