Wan 3.0 Reference Label Contract

Use this offline skill before a Wan 3.0 multimodal reference call. It maps redacted media roles to deterministic Image 1, Video 1, and Audio 1 prompt labels and blocks holes, overages, mixed first-frame/reference modes, and unclear authorization.

Input and output

Provide JSON with media, prompt_labels, duration, uses_boundary_frames, uses_file_or_link, and human_confirmed. The executor emits JSON and Markdown with a label map, blockers, and api_call_authorized=false.

Rules

• - Images are numbered separately from videos and audio. The first reference of each type is label 1.
• Maximums are 10 images, 5 videos, and 5 audio references. Video and audio reference duration totals may not exceed 15 seconds each.
• Boundary-frame inputs cannot be mixed with reference media. File and link inputs cannot be used together or with media references.
• Every declared label must occur in the prompt and every prompt label must map to exactly one declared reference.

Boundaries

This is a static label and budget checker. It does not fetch URLs, inspect media, call Wan, or prove that a prompt follows the intended subject. Rights, privacy, upload, cost, and release decisions stay with the authorized operator.