Video Model Runway
This Skill is the Runway model adapter layer. Before starting, follow the unified model adapter contract of the dependency video-generation; this file stores only Runway-specific differences. It is responsible solely for compiling already-defined material / SHOT.md creative intent into prompts that Runway's current model can execute more easily, and for verifying the actual input and duration capabilities of the selected Runway model; it does not redefine shot purpose, characters, plot, or product facts.
1. Fixed Specific Model and Generation Mode
First, clarify the actual Runway model and mode the user is using—for example, the current mainline Gen-4.5 for Text-to-Video or Image-to-Video.
If the user is using older Gen-4 / Gen-4 Turbo, Apps, Agent, Workflows, or other Runway tools, do not directly apply Gen-4.5's parameters; verify against current official documentation in references/MODEL-GUIDE.md.
Whenever precise duration, aspect ratio, frames per second (FPS), resolution, or input type is involved, always rely on the current official page.
2. Image to Video
When an input image already provides the subject, composition, lighting, and style, the prompt should primarily describe camera movement, subject motion, environmental motion, motion style, direction, speed, timing, and any changes that are genuinely needed for the current shot.
Do not repeat lengthy descriptions of static visual details already clearly expressed by the input image. Only add visual explanations when introducing new elements, obvious deformations, interactions, or changes to the starting frame.
Start with simple, direct motion descriptions by default, then add one major constraint based on failure patterns.
3. Text to Video
Without an input image, describe both visuals and motion simultaneously. A useful structure:
Cinematography / Composition → Subject → Action → Environment → Necessary Lighting / Style → Timing.
Not every item must appear. Preserve necessary creative freedom; avoid stacking contradictory long prompts just to be "complete."
4. Temporal Order
Complex actions may use natural language sequences or coarse timestamps. The time allocation must match the real time required by the action; do not break simple shots into overly dense instructions just to control every second.
5. Positive, Direct Descriptions
Prefer describing what should happen, rather than listing many negations. When a stationary camera is needed, write a clear positive state, e.g., "Lock the camera position; the camera remains still throughout the entire shot."
Hard constraints are still stored in SHOT.md; the final prompt retains only the expressions that the model truly needs to execute at that moment.
6. Return to video-generation
Output the current Runway prompt, suggested input image / reference, and parameters that need to be confirmed in the generation UI; long-term prompts are saved to the corresponding shot's prompt_vNN.md. When comparing multiple models in parallel, use clear model suffixes.
Packaging of one-shot generation bundles, file naming, and result importing are still handled by video-generation.
Completion Criteria
The prompt matches the specific Runway mode; Image-to-Video does not repeat descriptions of the input image, and Text-to-Video covers both necessary visuals and motion; precise capabilities are verified against current official documentation.