| name | grok-video-prompt |
| description | Use when converting shot-design results or storyboard first-frame beats into Grok-friendly prompts for `rhart-video-g/image-to-video`, `xai/grok-imagine/image-to-video`, or 全能视频G in the current flashxiaolagu mainline. This skill is assembly-only and enforces the tested Chinese segmented timeline format with `cut` and the fixed ending `无音乐,无音效,无字幕`. |
Grok Video Prompt
Overview
This skill converts shot-design.json into structured Grok image-to-video prompt data.
It is not a creative rewrite stage.
Its job is only:
- read shot-design
- keep one first-frame image per shot
- choose a legal Grok endpoint configuration
- translate the beat into the tested Grok prompt format
Authority
Follow:
${OPENCLAW_HOME:-~/.openclaw}/agents/flashxiaolagu制作/production-optimization-spec.md
- current project
analysis/shot-design.json
If there is a conflict, the current flash spec and the actual endpoint limits win.
Input
analysis/film-type-analysis.json
analysis/shot-design.json
prompts/banana-grid-shot-prompts.json
- split storyboard frame image paths
Output
prompts/grok-video-prompts.md
prompts/grok-video-prompts.json
Hard Endpoint Constraints
Current target endpoint:
rhart-video-g/image-to-video
Hard limits:
- only
image_to_video
- only one input image
duration: 6s / 10s / 15s
aspectRatio: 2:3 / 3:2 / 1:1
resolution: 720P / 1080P
If the source asks for an unsupported value, round to the nearest safe supported value and record why.
For current flash mainline portrait storyboard frames, prefer:
aspectRatio: 2:3
resolution: 720P
duration: 6s by default
Only move up to 10s or 15s when the shot clearly needs a longer speaking beat or a longer emotional turn.
Core Rules
1. Assembly only
Do not invent:
- new actions
- new dialogue
- new camera moves
- new characters
- new scene geography
2. Single-frame logic only
This route is first-frame-only.
- use
first_frame_image
- do not require
last_frame_image
- do not write prompts that depend on a far-away ending pose the first frame cannot support
If shot-design.json says the original shot was better suited for start-end generation, compress it into a smaller first-frame-driven motion beat instead of forcing a fake end-state.
3. Keep the user's tested Grok format
Every final prompt must use this structure:
根据参考图生成视频:
0-4 秒
[beat 1]
cut
4-6 秒
[beat 2]
无音乐,无音效,无字幕
Rules:
根据参考图生成视频: must be the first line
- each time range stays on its own line
- each beat follows directly under its time range
cut stays on its own line
- the final line stays exactly
无音乐,无音效,无字幕
4. Beat writing rules
Write in Chinese.
Each beat should usually be only 1-2 short sentences.
Prefer direct subject-action phrasing anchored to the visible first frame.
If there is spoken dialogue:
- write it as
角色 + 情绪 + 的说:台词
- preserve the spoken line itself
- do not translate Chinese dialogue into English
If the source line is already English, keep it in English.
If the shot has no dialogue, do not add any.
If the shot has both action and dialogue, keep the action brief and let the speech carry the beat.
5. Segment rhythm follows the selected duration
Recommended default splits:
6s: 0-4 -> cut -> 4-6
10s: 0-3 -> cut -> 3-7 -> cut -> 7-10
15s: 0-4 -> cut -> 4-8 -> cut -> 8-12 -> cut -> 12-15
You may adjust the boundaries slightly if the shot clearly needs a longer hold or a faster snap, but the final block must end exactly at the chosen duration.
6. Motion scale rule
Because Grok only gets one reference image, prefer:
- reaction
- gaze shift
- mouth movement
- hand movement
- shoulder tension
- small lean
- small forward or backward step
Avoid asking for:
- long-distance blocking travel
- large staging rewrites
- off-screen character entrances not supported by the first frame
- drastic costume or prop changes
When the source shot is mainly a spoken beat, let the dialogue carry the change instead of inventing large body motion.
7. Sound policy is fixed
Always end with:
Do not request:
- background music
- sound effects
- subtitles
- watermark text
Spoken dialogue is allowed only when it already exists in the source shot.
8. Source field priority
When translating a shot, use these fields in this order:
shot_prompt_cn
dialogue
first_frame_cn
last_frame_cn only as a weak motion hint
split_decision_note
video_mode_reason
last_frame_cn is never a hard ending contract for this skill.
9. Output contract
Each Grok item should include:
shot_id
mode
endpoint
first_frame_image
last_frame_image: null
duration_seconds
duration_label
aspect_ratio
resolution
prompt
prompt_segments
translation_note
Use:
mode: image_to_video
endpoint: rhart-video-g/image-to-video
Canonical Example
根据参考图生成视频:
0-4 秒
女孩娇嗔的说:你不记得了吗?上周的学术会议
cut
4-6 秒
女孩愤怒的说:你喝多了
无音乐,无音效,无字幕
Validation Checklist
- every shot uses
rhart-video-g/image-to-video
- every shot has exactly one
first_frame_image
- every shot keeps
last_frame_image: null
- duration is one of
6s / 10s / 15s
- aspect ratio is one of
2:3 / 3:2 / 1:1
- prompt starts with
根据参考图生成视频:
- every timeline block has valid start and end seconds
cut only appears between beat blocks
- the final line is exactly
无音乐,无音效,无字幕
- dialogue is preserved instead of being translated away
- no invented beat exceeds what the first frame can plausibly support