Omni Flash Video

Overview

Use this skill to turn a video idea plus reference media into a practical Gemini Omni Flash / Google Flow generation plan. Prefer short, atomic clips and clear reference roles over one overloaded prompt.

Default Workflow

1. Identify the generation mode:

- Frames to Video for a still image that should animate. - Ingredients/References to Video when preserving a character, object, style, voice, or environment across clips. - Video to Video editing when editing an uploaded/generated clip. - Storyboard / multi-shot when a sequence needs several generated clips.

1. Assign each input a role before prompting:

- start frame - end frame - character/identity reference - style reference - destination/environment reference - prior-shot continuity reference

1. Write one prompt per clip. Keep each clip to one main action, one camera

behavior, and one transition intent.

1. Generate one or more canaries in Flow. Review for first-frame match,

action readability, last-frame usability, and continuity.

1. If a clip is close, iterate conversationally in Flow with specific edits.

If the composition is wrong, revise the prompt/reference setup and regenerate.

1. Save prompts, reference paths, generated outputs, and review notes in the

project folder.

Read References

• Read references/official-research.md when you need current access paths,

supported Flow features, durations, limits, or source links.

• Read references/prompt-playbook.md before writing prompts, especially for

image-to-video, start/end frames, or door/portal transitions.

• Read references/paper-portal.md when working on the sketch-to-door project.

Prompt Rules

• Prompt motion from the reference rather than re-describing every visible

detail in the image.

• Use positive instructions: static camera, locked-off shot, `clean pencil

sketch animation, one continuous shot`.

• Keep style changes at transition points. Doorways/portals are useful reset

points for model consistency.

• Use short clip durations for controllability. Treat 4-10 seconds as a clip

unit, not a whole story.

• Avoid asking for several story beats in one generation. Split them.
• Use negative notes sparingly and as nouns/qualities: `extra limbs, extra

characters, colored background, clutter, text`.

Output Checklist

For each generated clip, record:

• Flow mode and model
• source frame/reference files
• prompt text
• duration and aspect ratio
• output file or URL
• what worked
• what failed
• whether the last frame is usable as the next seed