Generate Video

For a long recording cut into actual short clips, use repurpose-video. For standalone scripted narration, use generate-voiceover; an audio result does not establish that it was muxed into a video.

Create one clear motion asset using a compatible model, ready workspace references, and the current tool's nested parameters.

Workflow

1. Resolve scope with simplified-workspace when a workspace/teamspace is named or uncertain. Carry space_id into every related call.
2. Define the subject, action, placement, aspect ratio, duration, camera behavior, and audio needs. Infer ordinary choices from the user's brief; ask only for missing constraints that change the result or cost.
3. Discover current models with api_listVideoModels or api_getModelFields(type: "video"), depending on the exposed tools. Inspect api_getVideoModelFields or the available model-fields tool for model-specific durations, resolutions, and capabilities. Do not guess model IDs, costs, or supported values.
4. Use manage-assets to find/import/upload references and check readiness with api_getAsset. Every video input slot takes a Simplified asset UUID from the same workspace, not a local path or remote URL.
5. Choose storage: "asset" for a reusable video. Generation spends credits; proceed for an explicit request, and clarify material ambiguity in scope/cost before submitting.
6. Call api_generateVideo once. Current apikit waits for the rendered result; wait for this call before using any continuation tool.
7. Inspect the actual terminal status and return the rendered video link plus its real asset ID when present. Handle pending/timeout/older deployments as described in references/models-and-polling.md.

Current input contract

Top-level: model, parameters, and chosen storage; optional capability and space_id. All generation controls belong inside parameters.

Despite _url in the names, these slots take asset UUIDs. Use the same public slot names across models. Choose only controls the selected model supports, such as duration, resolution, aspect_ratio, or generate_audio. Image-driven modes may derive their ratio from the reference; do not force an unsupported ratio. Omit capability to let the current tool infer it from inputs, or use an explicitly supported mode.

{
  "model": "<discovered image-to-video model ID>",
  "storage": "asset",
  "parameters": {
    "image_url": "<ready image asset UUID>",
    "prompt": "A slow push toward the product, soft light moving across its packaging"
  }
}

Replace placeholders before calling. Do not assume every model supports every slot, duration, or audio option.

Creative quality

Use one readable visual beat for a short clip. Preserve product geometry, packaging, logos, and identity in references; avoid contradictory camera moves. Leave space for overlays when needed. If the client can inspect the output, check it; otherwise state that visual quality has not been inspected. Synthetic people/scenes are not documentary evidence.

Handoff

Use the client’s supported video preview or a usable link; preserve any thumbnail link and permanent asset UUID. Signed URLs can expire. default gallery storage is not a guarantee of a standalone reusable asset; transient may supply only a temporary URL. If no asset ID exists and the user wants reuse, import the downloadable output using api_createAsset. Never invent an asset ID from an art or variation ID.

Pass confirmed assets to simplified-social for drafts when requested. Generating a video does not authorize publishing. Report failures directly and avoid duplicate paid submissions after a timeout.