Generate Video

Generate Video

Create a production-appropriate video by discovering the live model contract, generating with valid parameters, and returning a reusable result.

Guardrails

• Video generation spends credits. Proceed when generation is explicit; confirm before spending when the user is exploring options or has not approved generation.
• Call api_getModelFields before generation. Never rely on a memorized model list, price, duration, aspect ratio, or parameter schema.
• Use only parameters returned for the selected model and capability.
• File-typed parameters take permanent Simplified asset UUIDs, not local paths, signed URLs, or arbitrary remote URLs.
• Use storage: "asset" when the video will be posted, reused, or retained. Use transient only for an explicitly temporary result.
• Do not treat the submission task_id as render completion. Follow the variation status contract in [references/models-and-polling.md](references/models-and-polling.md).
• Show returned URLs as clickable links; never embed them.

Workflow

1. Define the job: audience, objective, placement, aspect ratio, duration, visual subject, action, camera language, pacing, brand constraints, references, audio needs, and required delivery format. Infer common social dimensions only when the requested placement makes them unambiguous.
2. Call api_getModelFields with type: "video" and no model ID to discover current choices. Shortlist by supported capability, quality, speed, estimated time, and credit cost—not model fame alone.
3. Call api_getModelFields again with the selected model_id and capability to retrieve exact required fields, enums, defaults, and file-field expectations.
4. For image-to-video, multiple-image, or first/last-frame work, ensure every reference is a permanent asset UUID. Use the Simplified signed-upload asset workflow for user-supplied local files.
5. Write a motion prompt that describes subject, action over time, environment, camera behavior, composition, lighting, pacing, and exclusions. Avoid stacking contradictory movements or scene changes into a short clip.
6. Show the chosen model, capability, duration, format, credit information when available, storage mode, and prompt before generation when cost or creative direction remains ambiguous.
7. Call api_generateVideo with the exact nested parameters contract and the chosen storage mode.
8. If the call returns before terminal completion, preserve both returned IDs and call api_getVideoVariation until job_status is DONE or FAILED. Use a reasonable interval; do not busy-loop.
9. On success, report the reusable asset UUID, rendered video link, thumbnail link when available, model, format, and any material limitations. On failure, surface the provider error and suggest one targeted correction.

Creative Standard

• Design one clear visual beat per short clip. A six-second asset needs a readable action, not a miniature screenplay.
• Match the first frame and opening motion to the social hook; assume many viewers begin muted unless generated audio is central to the concept.
• Preserve product geometry, logos, packaging, people, and claims when references are supplied. Flag visible inconsistencies instead of presenting them as final.
• For paid or conversion creative, leave intentional visual space for on-screen copy and CTA overlays.
• Treat AI-generated people, testimonials, product behavior, and locations as synthetic; never imply documentary proof.

Output

Lead with the creative choice and generation status. Then provide the permanent asset ID, result links, intended placement, model/capability, and next useful action such as drafting a post or producing a controlled variant.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: celeryhq
• Source: celeryhq/simplified-ai
• License: MIT

Install and usage instructions live in the source repository linked above.