| name | dreamina-video |
| description | Use when generating commercial short videos with 即梦, Dreamina, Seedance, Seedance 视频生成, 种草视频, 带货视频, 获客视频, 推广视频, 图生视频, 多模态视频, or when a user provides image/audio/video references and asks for a stable marketing video. |
即梦 / Seedance 商业视频生成
Use this skill to turn references and a business goal into a stable short-video generation workflow. Creative judgment stays in the agent; provider calls, API keys, parameter validation, task status, downloads, storage, credits, and logs stay in MCP/server tools.
Required MCP Tools
| Tool | Use |
|---|
get_project_video_profile | Read project video defaults, model policy, model catalog, and credit multiplier before planning. |
register_video_reference | Upload or normalize a text/image/audio/video reference into an OSS-backed, Ark-accessible URL. |
validate_video_generation_params | Preflight model/ratio/resolution/duration/references and return estimated dynamic credits. |
build_video_generation_plan | Validate parameters and produce generation-plan.json plus SDK payload preview and estimated credits. |
create_video_generation_task | Submit the server-side Ark SDK async task. |
query_video_generation_task | Poll task status and retrieve generated URLs/metadata. |
download_video_generation_result | Download the completed provider result, upload it to OSS, and register it as a task file. |
Never call Volcengine/Dreamina HTTP APIs directly, never handle API keys, never use fixed global model/credit defaults, and never use the 即梦 CLI as the main execution path. Agents must use MCP tools. Do not treat external @-style labels as the server execution protocol; translate them into reference_role plus MCP references.
Workflow
- Work inside the existing project/plan/task flow. A video generation job is a
video task, not a separate product line.
- Call
get_project_video_profile(project_id) first. Use the returned project defaults, model policy, model catalog, and credit_multiplier; do not hardcode model IDs, default resolution, duration, or fixed credits.
- Prepare the workspace with
prepare_workspace(content_type="video", task_id=...) when available. Local files are temporary Claude workspace artifacts only; anything persistent must become an OSS-backed task file through MCP/server tools.
- Read the relevant references:
- Business structure:
references/methodology.md
- Consistency/retry rules:
references/stability.md
- Artifact and prompt templates:
references/prompt-templates.md
- MCP contract details:
references/mcp-contract.md
- Collect inputs: references for people/objects/scenes, business purpose (
planting, ecommerce, lead_gen, promotion), duration, ratio, resolution, model key, seed/camera preferences. Missing values come from project video profile, not from this skill.
- Register every non-text reference with
register_video_reference. Prefer task_file_id or upload through the platform; for video input, use the returned reference with server-measured input_duration_seconds and never hand-write a raw video_url into plan/create. Raw provider URLs are intermediate only and must not be the main delivery link.
- Write
reference-anchors.md: first declare reference_role for every reference, then separate each reference into must-keep, can-change, and must-not-change anchors.
- Write
script.md: use the 0-3s / 3-10s / 10-13s / 13-15s commercial structure. Do not submit raw marketing copy as a video prompt.
- Write
shot-plan.md: 4-5 shots for 15s by default. Each shot needs subject, action, scene, camera movement, visual focus, and negative constraints.
- Build the final prompt using global anchors + shot instructions + negative constraints. Keep one intent per shot.
- Call
validate_video_generation_params or build_video_generation_plan; save the result to generation-plan.json, including estimated dynamic credits and pricing breakdown. Show/record the estimate before submission.
Retry Rules
- Subject drift: strengthen
reference-anchors.md; reduce shot count.
- Product shape/color drift: move product descriptors into every product shot; avoid metaphorical phrasing.
- Action chaos: simplify action verbs; one action per shot.
- Scene jumping: repeat scene anchors and set
camera_fixed when appropriate.
- Too ad-like for planting: soften CTA and rewrite script as first-person experience.
- Weak conversion for ecommerce/lead gen: make benefit and CTA more explicit, but keep claims compliant.
- Unsupported model/parameter combo: only use server-returned validation errors or
suggested_params; do not invent a downgrade outside the project policy.
Record every retry in iteration-log.md with changed fields, reason, and result.
