# Seedance 2 Prompt

> Render a single 3-15s video clip via Seedance 2.0. Supports two backends: OpenRouter (default, model bytedance/seedance-2.0) and the official Volcengine ARK / BytePlus ModelArk endpoint (model doubao-seedance-2-0-260128 / dreamina-seedance-2-0-260128). Accepts a structured English video prompt, optional first-frame image, and optional identity/style reference image. Trigger when the user asks for AI video clip generation, 分镜视频, seedance, or wants a short cinematic shot from a prompt + frame.

- Skill: `gabrielmoreira/seedance-2-prompt` (Agent Skill)
- Install (CLI): `npx skillmds@latest add gabrielmoreira/seedance-2-prompt`
- Raw SKILL.md: https://api.skillmd.com/api/skills/gabrielmoreira/seedance-2-prompt/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: gabrielmoreira (https://skillmd.com/u/gabrielmoreira)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/gabrielmoreira/seedance-2-prompt

---


# seedance-2-prompt — Seedance 2.0 video clip generator (dual backend)

Submits a Seedance 2.0 generation job and downloads the resulting MP4.
Two backends share one CLI, picked via `with.provider`:

| `with.provider` | Endpoint | Auth env | Default model |
|---|---|---|---|
| `openrouter` (default) | `https://openrouter.ai/api/v1/videos` | `OPENROUTER_API_KEY` | `bytedance/seedance-2.0` |
| `volcengine` (CN) | `https://ark.cn-beijing.volces.com/api/v3/contents/generations/tasks` | `ARK_API_KEY` (or `VOLC_ARK_API_KEY`) | `doubao-seedance-2-0-260128` |
| `byteplus` (intl) | `https://ark.ap-southeast.bytepluses.com/api/v3/contents/generations/tasks` | `ARK_API_KEY` (or `BYTEPLUS_API_KEY`) | `dreamina-seedance-2-0-260128` |

Both flavours follow submit-then-poll, but differ in request shape
(OpenRouter uses a flat `prompt` field; ARK packs everything into a
`content[]` array), polling URL (OpenRouter returns `polling_url`; ARK
gets `id` and you construct `/contents/generations/tasks/{id}`), the
terminal-success status (`completed` vs `succeeded`), and where the
final MP4 URL sits (top-level `unsigned_urls[0]` vs `content.video_url`).
This script normalises both into a single Python contract.

## Inputs (`with:`)

| key | required | default | notes |
|---|---|---|---|
| `prompt` | yes | — | Structured English video prompt (use this skill's recipes). |
| `filename` | yes | — | Output `.mp4` path. |
| `provider` | no | `openrouter` | `openrouter`, `volcengine`, or `byteplus`. |
| `aspect_ratio` | no | `9:16` | `9:16`, `16:9`, `1:1`, `4:3`, `3:4`, `21:9`. |
| `duration` | no | `5` | Seconds, 3-15. OpenRouter Seedance accepts 4-15; a 3-second request generates a real 4-second provider clip and is trimmed locally to exactly 3 seconds. |
| `resolution` | no | `720p` | `480p`, `720p`, `1080p`. Ignored by OpenRouter. |
| `model` | no | provider default | Override model id. Empty means use provider default. |
| `input_image` | no | `""` | Strict first-frame path. If set, video starts from this image. |
| `input_reference` | no | `""` | Primary soft identity/style anchor path. Used only when `input_image` is empty. Same anchor passed across shots locks the character. |
| `input_reference_2` | no | `""` | Optional second reference (e.g. per-shot scene composition). Forwarded as a second `--input-reference` so the underlying provider sees both. Empty strings are filtered out before the API call. |
| `max_retries` | no | `0` | Extra retries for transient polling HTTP 429/5xx or transport failures after a job id is issued, capped at 5; every retry stays on that same job. A paid POST submit is never retried automatically, including an ambiguous 429/5xx or transport failure, because the provider may already have accepted and billed it. HTTP 401/403/other 4xx, terminal policy failures, and invalid downloaded media also stop immediately. `0` = no extra polling retry. |

**`input_image` vs `input_reference`** — `input_image` becomes the literal
first frame. `input_reference` is a softer style + identity hint the
model uses without locking the frame. For multi-shot drama, pass the
same `input_reference` to every shot; pass `input_image` only when you
want a specific opening frame.

## Prompt rules (from upstream + OpenSquilla tightening)

1. **One major action per 3-5s segment.** Don't pack multiple motions.
2. **Identity continuity** — repeat the main character's full
   description in every shot's prompt.
3. **Specific over poetic** — `"a young woman in a red trench coat
   walks through rain-soaked neon streets"` >> `"a woman walking"`.
4. **Negative constraints inline** — append `no watermark, no logo,
   no subtitles, no on-screen text.`
5. **IP-safe** — invent original character/brand names.
6. **Aspect ratio explicit** — append `aspect_ratio: 9:16`.

See `references/recipes.md`, `references/modes-and-recipes.md`,
`references/camera-and-styles.md` for the upstream playbook.

## Auth

- `openrouter` provider API-key resolution order:
  1. `--api-key` CLI argument
  2. The parent-injected atomic tuple
     `OPENSQUILLA_META_CAPABILITY_PROVIDER`,
     `OPENSQUILLA_META_CAPABILITY_API_KEY`, and
     `OPENSQUILLA_META_CAPABILITY_BASE_URL`; optional
     `OPENSQUILLA_META_CAPABILITY_PROXY` is used only for the matching
     provider's authenticated API requests
  3. `OPENSQUILLA_META_OPENROUTER_API_KEY`, retained for older parent runtimes
     and bound to OpenRouter's official origin
  4. `OPENROUTER_API_KEY` for direct CLI use, also bound to the official origin
- The generic parent key is accepted only with its matching provider and base
  URL. Parent and canonical environment credentials allow same-origin path
  changes through `--base-url`, but never a scheme, hostname, or effective-port
  change. Direct CLI users who intentionally select a different origin must
  pass both `--api-key` and `--base-url`.
- The child never discovers or parses `opensquilla.toml` from its workspace
  and never honors a workspace-selected arbitrary `llm.api_key_env`. Configure
  the active Gateway normally; its parent runtime performs that resolution.
- When a parent-resolved profile-pool credential receives an authentication,
  credit, or rate-limit failure, OpenSquilla parks that key for the next
  explicitly authorized run. It never repeats the current paid generation
  automatically.
- `volcengine` / `byteplus` provider reads `ARK_API_KEY` (with provider-
  specific fallbacks `VOLC_ARK_API_KEY` / `BYTEPLUS_API_KEY`). No
  config-file fallback for these — the OpenSquilla `[llm]` config
  describes the agent's selected LLM provider, not ARK / BytePlus video
  credentials.
- All three send the key as `Authorization: Bearer <key>` only to an API URL
  whose scheme, hostname, and effective port exactly match the configured
  provider base. Authenticated redirects are not followed. A provider-supplied
  cross-origin polling URL is ignored in favour of the canonical job URL under
  the trusted base. A parent-injected proxy is used for these matching provider
  API requests only and never for anonymous media downloads.
- Media downloads are always anonymous, including URLs hosted by OpenRouter.
  Download URLs and every redirect must use HTTPS, contain no userinfo, and
  resolve exclusively to public addresses; localhost, private, loopback,
  link-local, reserved, unspecified, and multicast targets are rejected. Each
  connection is pinned to the address that passed validation while retaining
  the original TLS SNI/hostname, eliminating DNS-rebinding between validation
  and connection.

## Output

Prints the absolute path of the saved `.mp4`, then a
`VIDEO_GENERATION_RECEIPT: {...}` line on stdout. A matching sanitized
receipt is saved as `<filename>.receipt.json`; it records provider,
model, safe job id, and ffprobe verification metadata for both the
provider clip and final clip. Downloads first land in a private
same-directory candidate file. The script requires a real video stream,
positive duration, reasonable dimensions, and duration within tolerance;
3-second requests are probed both before and after the local trim. Only a
verified candidate is atomically published with `os.replace`, so an HTML
error page, truncated download, or failed trim cannot overwrite an existing
output. The receipt never contains the API key, prompt, raw provider response,
request/generation metadata, usage payload, or signed download URL. Logs omit
URL queries/fragments and raw provider messages. Non-zero exit on any error;
stderr carries a sanitized status/code summary when available.

On a recognized provider-policy rejection, the script stops immediately and
persists a failure sidecar instead of fabricating API success. That sidecar is
strictly limited to `status=policy_rejected`, provider/model,
`reason=provider_policy_rejected`, and the allowlisted `policy_code`; it never
contains raw provider text, signed URLs, request/job IDs, prompts, or secrets.
Meta workflows can then run their local fallback while preserving an honest,
user-visible degraded reason for the delivery audit.

## Cost / latency

- OpenRouter `bytedance/seedance-2.0` 5s @ 9:16 720p: ≈30-90s wall, ≈$0.76.
- Volcengine official 5s 1080p: ≈30-120s wall, ≈$0.93.
- Volcengine `doubao-seedance-2-0-fast-260128`: roughly half cost and faster.
- The returned `unsigned_urls` / `content.video_url` expire 24 hours
  after success on the Volcengine path — this script downloads them
  before that window so the local mp4 is durable.

## Multi-segment workflows (>15s)

Generate segments individually with `duration ≤ 15`, ending each on a
stable hand-off frame. Stitch with the `video-merger` skill. See
`references/modes-and-recipes.md` § "Multi-Segment Stitching".