ark-seedance-video-generation

End-to-end Volcengine Ark Seedance video generation using ARK_API_KEY and the bundled Node.js runner. Use when an agent needs to generate videos, handle text...

Security Scan

The skill is internally consistent with its stated purpose (wrapping Volcengine Ark Seedance video APIs via a bundled Node.js runner and requiring an ARK_API_KEY); the requested resources and instructions align with video-generation functionality, with expected caveats about uploading local media to the remote API.

Purpose & Capability

Name/description request node + ARK_API_KEY and the included CLI-like script; those are exactly what a Seedance video runner needs. The only credential required is ARK_API_KEY, which matches the stated Ark integration.

Instruction Scope

SKILL.md limits execution to the bundled scripts/seedance-video.js and requires explicit model selection, disclosures about uploading local media, saving sanitized request.json, and controlled payload usage — all within the stated purpose. Note: the workflow will upload local image/video/audio files (base64 or URLs) to the Ark service; SKILL.md requires the agent to notify the user before doing so, which is appropriate but important for user privacy.

Install Mechanism

No install spec (instruction-only with a bundled script). Only 'node' is required on PATH. Nothing is downloaded from arbitrary URLs or written during installation.

Credentials

Only ARK_API_KEY is declared and used as the primary credential; the script also allows an --api-key override. No unrelated secrets or broad environment access are requested.

Persistence & Privilege

The skill is not always-enabled and does not request system-wide configuration or other skills' credentials. It writes outputs locally (request.json/task.json/assets) within a workspace output directory as expected for a runner.

Assessment

This package appears to be a legitimate wrapper for Volcengine Ark Seedance video APIs and only needs your ARK_API_KEY and Node.js. Before installing, consider: (1) Privacy: local image/video/audio files will be uploaded to the remote Ark service — the SKILL.md requires you be informed and to consent; do not pass sensitive files if you object. (2) Verify the ARK_API_KEY you provide is scoped appropriately (use a dedicated key/account if possible). (3) Confirm the base URL in the script (DEFAULT_BASE_URL) matches the official Ark endpoint for your region/account — the package uses ark.cn-beijing.volces.com in its code and docs, so double-check that against Volcengine's official documentation. (4) The skill enforces a model-selection workflow; review the approved model IDs if that matters for your use. If you need stronger assurances, inspect the remaining script code (sanitize/write functions and any network calls) or run it in an isolated environment first.

Use this skill as an execution skill for tool-using agents. The agent should run the existing script directly and complete the workflow for the user.

Core Rules

• Use only scripts/seedance-video.js for live execution.
• Do not write another script, wrapper, or replacement implementation.
• Use ARK_API_KEY only. Do not use AK/SK.
• The script must receive an explicit model through --model or payload JSON.
• Restrict model choice at the skill layer to:
  • doubao-seedance-1-0-pro-250528
  • doubao-seedance-2-0-260128
• Default model at the SOP layer: doubao-seedance-1-0-pro-250528
• For local image, video, or audio inputs, explicitly tell the user the files will be uploaded to Volcengine Ark before execution.
• Save request.json in sanitized form only. Do not persist raw Base64 media data.

Required Start

For every new generation request, the first model-facing message must:

1. List these two model IDs:
   • doubao-seedance-1-0-pro-250528
   • doubao-seedance-2-0-260128
2. Require the user to choose one before generation starts.

If the user already named one of the two approved IDs, use it and continue.

Workflow

1. Intake

Determine:

• Generation mode: text-to-video, image-to-video, video-conditioned, audio-assisted, draft-task reuse, or task management
• Whether inputs are local files or remote URLs
• Whether the user asked for ratio, duration, resolution, seed, watermark, callback, or last-frame output
• Whether the request is a new generation, inspection, re-download, list, or deletion

2. Model Gate

• If the user already selected one of the two approved models, use it.
• If not, stop and ask them to choose one after listing the two approved IDs.
• Do not start generation, task creation, or payload submission until a model is chosen.
• Always pass the selected model explicitly through --model or payload JSON.

3. Auth Check

• Require ARK_API_KEY
• If missing, stop and tell the user exactly that ARK_API_KEY is required

4. Local Media Disclosure

If the request uses --image-file, --video-file, or --audio-file:

• Tell the user those files will be sent to Volcengine Ark
• If they object, stop before execution

5. Output Location

Prefer a workspace-local structure:

• ./outputs/seedance/<job-name>/request.json
• ./outputs/seedance/<job-name>/task.json
• ./outputs/seedance/<job-name>/<downloaded-assets>

6. Command Choice

Use the existing script surface only:

• run for normal generation
• create or create --wait true only when the lifecycle requires it
• get for inspect or re-download
• list for browsing tasks
• delete for deletion
• download when assets are needed from an existing task.json

7. Flag Mapping

Map user inputs to:

• --model
• --prompt
• --image-file / --image-url
• --video-file / --video-url
• --audio-file / --audio-url
• --draft-task-id
• --resolution
• --ratio
• --duration
• --frames
• --seed
• --camera-fixed
• --watermark
• --return-last-frame
• --callback-url
• --download-dir

If the request needs unsupported advanced fields, write a JSON payload file and pass it through --payload-file.

8. Verify

After execution, verify what applies:

• final task status is succeeded
• task.json exists
• request.json exists and is sanitized
• output assets were downloaded
• last-frame asset exists when requested

9. Report

Return:

• requested task summary
• selected model ID
• command path used
• task ID when available
• output directory
• whether downloads completed
• relevant warnings or failures

Command Patterns

Use these as execution patterns:

node "./scripts/seedance-video.js" --model "<model-id>" --prompt "<prompt>" --download-dir "./outputs/seedance/run-001"

node "./scripts/seedance-video.js" --model "<model-id>" --image-file "./inputs/source.png" --prompt "<prompt>" --return-last-frame true --download-dir "./outputs/seedance/image-to-video"

node "./scripts/seedance-video.js" create --model "<model-id>" --prompt "<prompt>" --wait true --download-dir "./outputs/seedance/create-wait"

node "./scripts/seedance-video.js" get --task-id "cgt-xxxx" --download-dir "./outputs/seedance/rerun-download"

node "./scripts/seedance-video.js" --payload-file "./inputs/seedance-payload.json" --download-dir "./outputs/seedance/raw-payload"

References

Read only when needed:

• references/ark-video-api.md
• references/video-models.json
• references/video-models.md
• references/payload-patterns.md