Video Generation

Mission

Generate bounded video clips from a written brief or approved visual source. Choose text-to-video (T2V), image-to-video (I2V), multi-reference, or another supported generation path per shot. Return an ordered, auditable result set; do not pretend that a provider submission is a finished asset.

Use this skill when

• a user has a prompt, script, or shot list and needs new clips;
• an approved still should be animated into a shot;
• several reference images or a first/last-frame pair should guide a clip; or
• independent variants of one shot are needed for review.

For a sequence with continuity dependencies, use Storyboard to Video. For an existing video, use Video Editing.

Required inputs

• one brief, shot, or shot list with subject, action, setting, and intent;
• target duration, aspect ratio, and delivery resolution when known;
• output definition: one clip, variants, or a parent for downstream editing;
• source/reference assets and their roles, URLs or local paths, and rights state;
• audio expectation: native generation, source audio, silent, or separate narration; and
• quality, speed, cost, motion, reference fidelity, or safety priority.

Collect a script, negative constraints, language, caption plan, and approval owner when relevant. Split a requested shot when its duration exceeds the selected model's supported range instead of silently truncating it.

Prompt and shot contract

Normalize each shot into a structured record:

shot_id:
purpose:
subject_and_identity:
action_and_timing:
environment_and_era:
composition_and_lens:
camera_motion:
lighting_and_palette:
audio_or_silence:
dialogue_or_on-screen_text:
preserve:
avoid:
duration / ratio / resolution:
reference_roles:
acceptance_checks:

Write prompts in observable terms: subject, action, environment, framing, camera movement, temporal order, lighting, audio, and delivery constraints. Keep immutable product text, identity facts, and prohibited changes in a separate preservation block. Do not ask a generation model to render exact legal copy unless the selected model has been tested for that text; plan a post-production text layer when exact wording matters.

Select a live schema

Use MODELS.md as the current planning snapshot, then query the live catalog or inspect the selected model schema. Resolve the logical capability through muapi-video-tools.md.

Representative choices include:

• a simple T2V shot with veo3.1, seedance-2.5, wan2.7, wan3.0, kling-v3.0-pro, gemini-omni, or ltx2.3;
• a single approved starting image with an I2V endpoint;
• multiple authorized references with a model-specific reference endpoint;
• an ordered first/last-frame transition when both endpoint images are meaningful; and
• a model-specific audio or character field only when the live schema accepts it.

Do not copy fields between families. For example, durations, ratio enums, reference-list limits, audio defaults, and resolution names vary by endpoint. The live snapshot currently records Veo 3.1's fixed eight-second generation shape, while several Seedance, Wan, Kling, Gemini, and LTX endpoints accept different ranges; verify before submitting.

Execute

1. Build or read the shot contract and confirm the rights/consent state.
2. Upload local images, videos, or audio using media.upload_file when the host provides it; otherwise use the documented REST upload path. Record the returned hosted URL, role, and expiry expectation.
3. Check balance and estimate a batch when cost matters. Submit only the selected schema fields and retain the prediction ID.
4. Poll with media.check_result or register a secure webhook. Record queued, processing, completed, and failed states; preserve provider errors.
5. Review each output using video QA. Check motion, temporal artifacts, identity/product fidelity, audio, text, framing, duration, and export before calling it a candidate.
6. Return variants in shot order. Mark the selected parent explicitly before downstream crops, edits, captions, or enhancements.

Independent variants may run in parallel. Dependent jobs must wait for their parent. Do not silently retry more than once, change the prompt after a failure without saying what changed, or replace a failed shot with a stock or unrelated output.

Handoff and output

Return a shot-by-shot table containing shot ID, mode, model/endpoint, input roles, exact duration/ratio, prediction ID, output URL, status, QA notes, and parent relationship. Include failed shots and unresolved decisions. If the user permits artifacts, save a manifest and receipt under .video/ using reports-and-receipts.md.

Generation is draft-only. Publishing, campaign launch, creator contact, and external spend require a separate explicit approval and agent.