/reference-driven-video-prompting

name: reference-driven-video-prompting
description: Use for image-to-video, reference-video continuity, keyframe-driven prompting, camera control, and consistency workflows based on official Runway, Kling, and character/style reference patterns.
tags:
  - image-to-video
  - reference-video
  - keyframes
  - camera-control
  - consistency
agents:
  allow:
    - planner
    - image_designer
    - image_edit_agent
    - video_designer
    - flf_video_designer

Reference Driven Video Prompting

Use this skill when the workflow depends on image inputs, prior videos, keyframes, or reference-based consistency.

Source Basis

• Runway Gen-3 prompting, keyframes, and camera control docs.
• Kling image-to-video prompt pattern.
• Midjourney character/style reference concepts for identity/style persistence.

Reference Hierarchy

1. identity reference 2. world/location reference 3. previous formal shot continuity 4. style reference 5. motion prompt

Prompting Rules

• With an input image, focus text on desired motion and camera behavior rather than re-describing the full image.
• Use direct, descriptive motion language; avoid conceptual chatter.
• Separate subject motion from camera motion.
• When using keyframes, define what must transform between first/middle/last frame.

Consistency Rules

• Character references preserve face, hair, clothing silhouette, and recurring identity traits.
• Style references preserve palette, texture, light behavior, and rendering attitude.
• Previous video references preserve action phase, movement direction, and world continuity.

Camera Control Rules

• Choose one dominant camera move per shot unless complexity is essential.
• If the subject moves toward camera, note whether the camera compensates, holds, or retreats.
• Do not overload a shot with contradictory camera verbs.

Good Motion Language

• sways gently in wind
• advances in heavy armor with restricted weight
• camera slowly pushes in as expression hardens
• orbit tightens during transformation

Avoid

• vague commands with no motion target
• multiple unrelated action verbs in one short shot
• re-randomizing identity when continuity references already exist