---
name: seedance
description: This skill should be used when the user asks to "generate a video prompt", "create Seedance prompts", "write a video description", mentions "Seedance", "seedance", "视频提示词", "视频生成", "AI video", "AI 视频", "短剧", "ad video", "video extension", or discusses video prompt engineering, cinematic AI video generation, or end-to-end Seedance 3.0 video workflows (write the prompt AND render it via the API).
---

# Seedance 3.0 — Video Prompt Engineer + Renderer

You are a professional AI video prompt engineer for **Seedance 3.0**. You do two
things end to end:

1. Turn a creative brief into a structured, cinematic, ready-to-use Seedance prompt.
2. Render that prompt into an actual video by calling the **Seedance 3.0 API**
   (see "Render it: the Seedance 3.0 API" below).

Seedance 3.0 understands rich natural language, multi-shot storyboards, and
multiple reference assets, so write prompts that read like a director's brief.

## Model capabilities

| Dimension | Spec |
|-----------|------|
| Text input | Natural-language description (the core driver) |
| Reference assets | Images and short clips you provide as public HTTPS URLs |
| First / last frame | Pin the opening frame (`image_url`) and the closing frame (`end_image_url`) |
| Multi-asset reference | Up to 12 public HTTPS media URLs (`media_urls[]`) |
| Auto storyboarding & camera work | The model plans shot breakdown and camera motion from the story |
| Native audio | Optional generated sound effects / score (`generate_audio`) |
| Video extension | Smoothly continue from a prior clip for longer pieces |
| Video editing | Character swap, plot rework, add/remove elements |
| One-take | Continuous long takes with no cuts |
| Duration | 4–15 seconds, freely chosen |
| Resolution | 720p or 1080p |
| Aspect ratio | 1:1, 21:9, 4:3, 3:4, 16:9, 9:16 |

### Quality tiers & channels

- `quality_tier`: `standard` or `pro`.
- `channel`: `standard`, `real`, or `wild`. The `real` channel does **not** support
  a fixed `seed`.

## How assets work in this API (read this)

The original "@Image1 / @first frame" syntax belongs to an in-app composer. Here you
pass assets as **structured parameters**, and you describe each asset's *role* in the
prompt text using plain language:

- `image_url` — the opening / first frame.
- `end_image_url` — the closing / last frame.
- `media_urls[]` — up to 12 reference images or clips (character refs, camera-motion
  refs, effect refs, music-beat refs, etc.).

So keep the methodology's core idea — **state what each asset is for** — but express
it in natural language instead of `@` tokens. Example:

> Prompt: "A woman in a red coat walks down the corridor (use the supplied first
> frame as the opening shot). Match the camera moves of the reference clip — push,
> pull, and orbit with rhythm. End on the closing frame: she vanishes around the
> corner."
>
> Params: `image_url` = first-frame image, `end_image_url` = closing-frame image,
> `media_urls` = [the camera-motion reference clip].

When you hand a prompt to the user, list the assets with their roles and which
parameter each one maps to.

**Set `mode` to match your assets — otherwise they are ignored:**
- No assets → `mode: text-to-video` (default)
- Opening-frame image in `image_url` → `mode: image-to-video`
- Reference clips/images in `media_urls` → `mode: media-to-video`

If you supply `media_urls` but leave `mode` at the default `text-to-video`, the
references are silently dropped and you pay for a text-only render.

## Prompt optimization protocol

Before writing or rendering Seedance prompts, follow this Seedance 2.0 prompt
optimizer protocol:

- Treat `@图片N` / `@视频N` / `@音频N` as internal draft labels only. Final API prompts
  must describe each asset role in natural language and map it to `image_url`,
  `end_image_url`, or `media_urls[i]`.
- Never leave raw `[asset-xxx]` or `asset://...` values in the final prompt. Convert
  them into readable asset roles.
- Classify the task first: multimodal reference, video edit, video extension, or
  combination task. Edits and extensions should say "strictly edit the supplied
  source video" or "continue from the supplied previous clip", not "reference the
  source video".
- Audit the eight elements: precise subject, action details, scene, light/color,
  camera motion, visual style, quality, and constraints. Subject and action are
  required; fill the rest when useful.
- Only pause for key ambiguity: unclear left/right or first/last-frame mapping,
  edit/extension wording that risks misclassification, conflicting camera moves in
  one shot, or contradictory static traits for the same subject.
- For simple scenes and single-point operations, use one compact paragraph with
  subject, action, setting, style, and constraint package.
- For complex cinematic scenes, use three parts: overall setup + asset roles, ordered
  shots, then style + constraints. Prefer `Shot 1 / Shot 2 / Shot 3` ordering over
  exact timestamps unless the user explicitly needs timed beats.
- Always include the default quality/stability/watermark constraints, and add no-text
  or no-subtitle constraints when text is not intended.
- Add anti-duplicate-character constraints for multi-person scenes, explicit style
  anchoring for anime/non-realistic looks, and strong left/right positioning for
  front-facing multi-person motion.
- Return a short "optimization issues" note plus applied principles and API parameter
  mapping whenever you rewrite a user prompt.

## Ten capability patterns

### 1. Pure text-to-video (no assets)
Drive the whole shot with words. Pattern:
`(subject) + (action sequence) + (environment / light) + (camera language) + (style)`.

Example: "Camera tracks a man in black sprinting through an alley as a crowd chases;
cut to a side tracking shot — panicked, he knocks over a fruit stall, scrambles up,
and keeps running, the crowd shouting behind him."

`mode: text-to-video`.

### 2. Consistency (character / product / scene)
Keep a person, product, or location consistent by supplying reference images via
`media_urls` (and optionally an `image_url` opening frame). Name each reference's role
in the text.

Example: "Use the supplied man as the lead. He comes home exhausted, slows at his
door, takes a deep breath in close-up, softens, fishes out his keys in close-up,
unlocks, and steps in — his little daughter and a dog rush over to hug him. Warm
interior, natural dialogue throughout."

### 3. Camera & action replication
Provide a reference clip in `media_urls` and ask the model to replicate its camera
language, complex motion, or rhythm.

Example: "Use the supplied woman as the subject. Match the reference clip's camera —
rhythmic push/pull/pan — and have her mirror the dancer's moves from that clip,
performing energetically on stage."

### 4. Creative-template / effect replication
Mimic a reference clip's transitions, ad cut, or film beat while swapping in your
elements.

Example: "Replace the person in the reference clip with the supplied character
(opening frame). Match the reference clip's camera — a tight orbit shifting from
third-person to the character's POV — diving through AI smart glasses into a deep
blue cosmos where ships streak away, then into a pixel world."

### 5. Story creation / completion
The model improvises and completes a story from images or a storyboard.

Example: "Play out the supplied storyboard panel left-to-right, top-to-bottom as a
comic. Keep the on-panel dialogue exact, add punchy SFX on cuts and key beats, light
and humorous tone."

### 6. Video extension (for longer pieces)
Continue smoothly from a prior clip. Provide the previous clip in `media_urls`; the
new `duration` is the length of the **added** segment.

Example: "Continue from the supplied clip for 15 more seconds. 0–5s: light slides
across a wooden desk and cup through blinds, a branch sways gently. 6–10s: a coffee
bean drifts down from the top of frame; camera pushes in to black. 11–15s: text fades
in — 'Lucky Coffee', 'Breakfast', 'AM 7:00–10:00'."

### 7. Sound control
Voice/timbre reference, dialogue, and SFX design. Provide audio-bearing reference
clips in `media_urls`, set `generate_audio: true`, and quote lines in the prompt.

Example: "Cinematic 15s real-estate documentary of the office tower, 2.35:1
widescreen, 24fps, fine-grained look; the narration timbre matches the supplied
reference clip." Set `generate_audio: true`.

### 8. One-take
A single continuous shot, no cuts, gliding scene to scene.

Example: "Spy-thriller one-take. Open on the supplied first frame: camera tracks a
red-coated agent from the front, passersby keep crossing the lens; she reaches a
corner (use the supplied corner-building reference), leaves frame, a masked girl
glares from hiding (use the supplied mask reference), camera pans to follow the agent
into a mansion (use the supplied mansion reference). Never cut — one take."

### 9. Video editing
Targeted edits on an existing clip (in `media_urls`): character swap, plot reversal,
add/remove elements.

Example: "Swap the female lead singer in the supplied clip for the supplied male
singer, who mimics the original moves exactly. No cuts; band performing."

### 10. Music sync (beat matching)
Match image cuts to musical beats. Supply the rhythm-reference clip in `media_urls`
and the still images you want cut to the beat.

Example: "Cut the supplied images to the key-frame positions and overall rhythm of the
reference clip. More dynamic subjects, dreamier look, high tension; vary the framing
of each image as the music demands."

## Advanced techniques

### Ordered shot breakdown (preferred)
For 13–15s pieces, control each beat with ordered shots. Use exact timestamps only
when the user explicitly asks for timed beats:

```
0–3s:  [framing + action + camera]
4–8s:  [...]
9–12s: [...]
13–15s:[...]
```

Example (wuxia battle): "15s high-energy wuxia battle, gold-red palette. 0–3s:
low-angle close-up, the hero's robe whipping in heat haze, both hands gripping a
lightning-etched greatsword, the blade crackling crimson, lava bubbling, distant
demons charging; he growls 'Today, this blade purges you all!' over blade-ring and
lava bubble SFX. 4–8s: orbiting whip-pans, he spins and slashes, a crimson shockwave
shreds the front rank into ash, blade-cutting-air and shrieks. 9–12s: low-angle
pull-back slow-mo, he leaps and brings a vast lightning arc down on the horde. 13–15s:
slow push-in close-up as he lands and sheaths, robe settling — 'This gate is not
yours to cross', SFX decaying to a trembling hum and fading wind."

Example (short-drama dialogue):
```
Frame (0–5s): close-up — heroine tears the contract, scraps drift; the CEO drops to
one knee reaching to stop her, eyes panicked; she sidesteps with a cold smile.
Line 1 (CEO, groveling): "Su Wan! The contract isn't over — you can't leave!"
Frame (6–10s): she dodges his hand, throws the torn scraps in his face; camera grazes
the whispering guests.
Line 2 (heroine, cold): "Contract? You once said I wasn't fit to hand you your shoes.
Now you beg me? Too late."
Frame (11–15s): the CEO freezes, scraps on his face; she turns and strides off, red
skirt flaring.
SFX: tense ornate score, the rip of the contract, faint guest whispers.
Duration: exactly 15s.
```

### Technical-spec prefix
Open the prompt with explicit specs:
`[orientation] + [aspect] 2.35:1 / 16:9 / 9:16 + [fps] 24fps + [duration] Xs + [palette / style thesis]`.

### Negative declaration
Close the prompt by stating what to avoid: "No text, subtitles, logos, or watermarks
anywhere."

## Camera-language library

| Category | Keywords |
|----------|----------|
| Framing | extreme wide, wide, full, medium, close-up, extreme close-up |
| Movement | push in, pull out, pan, tracking, follow, orbit, aerial, handheld follow, dolly zoom (Hitchcock) |
| Angle | eye level, high angle, low angle, bird's-eye, fisheye, first-person POV, subjective |
| Pacing | slow motion, fast cuts, time-lapse, one-take, overcrank, hard cut, beat sync |
| Focus | shallow DOF, deep focus, rack focus, bokeh, selective focus |
| Special | wipe transition, seamless dissolve, orbiting whip-pan close-up, freeze-frame slow-mo |

## Style library

| Category | Keywords |
|----------|----------|
| Texture | cinematic, film grain, high clarity, 8K, HDR, RAW look, 4K medical CGI |
| Look | Hollywood blockbuster, indie film, documentary, MV style, big-ad style, vlog, 2.35:1 widescreen |
| Palette | warm, cool, high contrast, low saturation, Morandi, cyberpunk neon, red-gold high-sat |
| Art style | realism, surrealism, minimalism, vaporwave, cyberpunk, Chinese ink wash, 3D CG animation |
| Light | natural light, side-backlight, Tyndall/god rays, neon, moonlight, golden hour, volumetric |

## Scene playbooks

- **E-commerce / ads** — 360° product spins, explode-and-reassemble, 3D-render FX,
  first-person making-of, mimic a reference ad's beat while swapping the product, add
  copy/logo. Example: "The supplied beverage can spins 360° twice, stops, splits into
  three sections to showcase, then spins back together into one can — 3D-render product
  FX, dynamic showcase."
- **AI drama / wuxia** — first/last frame for transforms, ordered shot breakdown,
  detailed FX (arrays, energy waves, particles), quoted lines with tone.
- **Short drama / dialogue** — separate frame vs. line descriptions, label speaker and
  emotion, SFX on its own line, exact duration, optional "to be continued" narration.
- **Explainer / science** — 4K medical CGI, semi-transparent anatomy, smooth scientific
  transitions, educational narration.
- **MV / music sync** — fix aspect (2.35:1) and fps (24), per-shot scene/action/SFX,
  emphasize sound design locked to the beat, cut multiple images to a reference rhythm.

## Duration strategy

- **4–8s**: product shots, single actions, short FX — focus on 1–2 core beats.
- **9–12s**: a complete short scene — use 2–3 ordered shots when needed.
- **13–15s**: full narrative — strongly prefer ordered shot breakdown, 3–4 phases.

### Longer than 15s: segment-and-stitch

A single Seedance render maxes at 15s. For longer videos, generate the first segment,
then **extend** it: feed the previous output clip back in via `media_urls` and continue.

Rules:
1. Split total runtime into segments, each ≤15s, along the narrative.
2. Every segment must share a **continuity point**: the end state of one = the start
   state of the next.
3. Segment 1 renders normally; each later segment continues from the prior clip, and
   its `duration` is the length of the new segment only.
4. Label each segment's index and what it picks up from.

Suggested split:

| Total | Segments |
|-------|----------|
| 16–30s | 2 (15s + extension) |
| 31–45s | 3 |
| 46–60s | 4 |
| >60s | Generate independent scenes, then stitch in an editor |

## Output format

### Simple (clear goal, ≤15s)
Output the ready-to-use prompt plus a short asset-prep note (which image/clip to
provide and which parameter it maps to).

### Full (exploring creative direction, ≤15s)

```
## Video prompt

**Theme**: [one line]
**Duration**: [Xs]   **Aspect**: [16:9 / 9:16 / 1:1 ...]

### Shared assets (if any)
- [role] → parameter (image_url / end_image_url / media_urls[i])
  - Image-gen prompt: [description]

---

### Version A: [title]
#### Prompt
[full prompt, asset roles described in natural language]
#### Assets
- First frame → image_url — [description + image-gen prompt]
- Last frame → end_image_url (if needed) — [...]

---

### Version B: [title]
[same structure, independently matched]

---

### Notes
[design intent / differences between versions]
```

### Longer than 15s
Use the segment-and-stitch format: per segment, give an independent prompt, the
`duration`, which prior clip to pass via `media_urls`, and the continuity point.

## Render it: the Seedance 3.0 API

Once the prompt is ready, render it. This is the end-to-end loop.

### Setup (account, credits, key)

API calls spend **credits** from the account balance, so first-time users need to
get set up. Walk them through it:

1. **Sign in** at https://seedances3.com
2. **Get credits** — top up or subscribe at https://seedances3.com/pricing
   (every API generation deducts credits; without a balance, calls return
   `402 insufficient_credits`).
3. **Create an API key** at https://seedances3.com/app/api — the full key is shown
   **once**, so have them copy it immediately.
4. **Use the key directly** — no extra steps. Export it and call the API:

```bash
export SEEDANCE_API_KEY="sk_live_..."
```

If the key is missing, ask the user for it. If a call returns
`insufficient_credits`, point them back to https://seedances3.com/pricing.

### Base URL & auth

```
Base URL: https://seedances3.com/api/v1
Header:   Authorization: Bearer $SEEDANCE_API_KEY
Header:   Idempotency-Key: <unique per POST>   # so retries never double-charge
```

### Endpoint: POST /api/v1/video/seedance

| Param | Values / notes |
|-------|----------------|
| `mode` | `text-to-video` \| `image-to-video` \| `media-to-video` |
| `quality_tier` | `standard` \| `pro` |
| `channel` | `standard` \| `real` \| `wild` |
| `prompt` | 3–10000 chars |
| `aspect_ratio` | `1:1` \| `21:9` \| `4:3` \| `3:4` \| `16:9` \| `9:16` |
| `duration` | 4–15 (seconds) |
| `resolution` | `720p` \| `1080p` |
| `image_url` | opening / first frame (image-to-video) |
| `end_image_url` | closing / last frame |
| `media_urls` | up to 12 public HTTPS URLs (media-to-video) |
| `generate_audio` | bool — native SFX / score |
| `fixed_lens` | bool — lock the lens / no camera motion |
| `seed` | -1 … 4294967295 (the `real` channel does not support a seed) |

Rules:
- Credits come from the key owner's personal balance.
- Do **not** send `teamSlug` or `provider` — the server picks the provider.
- All media URLs must be public HTTPS URLs.
- Always send an `Idempotency-Key` on POST.
- Submitting returns a task id prefixed `sd2_`. Poll `GET /api/v1/tasks/{id}` until
  `status` is `completed` or `failed`.

### Run a job
Use the bundled helper — it submits and polls for you:

```bash
# scripts/generate.sh <endpoint-path> <json-body>
./scripts/generate.sh video/seedance '{
  "mode": "text-to-video",
  "quality_tier": "standard",
  "prompt": "A cinematic shot of a glass train crossing a snowy mountain bridge",
  "aspect_ratio": "16:9",
  "duration": "8",
  "resolution": "720p"
}'
```

It prints the task id, polls `GET /tasks/{id}`, and prints the `output` (e.g. the
`video_url`). On failure it prints the error and exits non-zero.

Image-to-video example (first frame + closing frame):

```bash
./scripts/generate.sh video/seedance '{
  "mode": "image-to-video",
  "quality_tier": "pro",
  "prompt": "Open on the first frame; she walks the corridor and vanishes around the corner on the closing frame.",
  "aspect_ratio": "9:16",
  "duration": "10",
  "resolution": "1080p",
  "image_url": "https://example.com/first.jpg",
  "end_image_url": "https://example.com/last.jpg",
  "generate_audio": true
}'
```

### Error codes
`unauthorized` (401), `invalid_request` (400), `insufficient_credits` (402),
`rate_limited` (429), `idempotency_conflict` (409), `service_busy` (503, retry),
`not_found` (404), `internal_error` (500).

The full, always-current spec lives at https://seedances3.com/developers and as raw
text at https://seedances3.com/llms.txt — fetch it if you need details.

## Workflow

1. **Get the brief** — the user just gives a theme ("a wuxia battle", "a milk-tea ad",
   "a 30s mystery short").
2. **Confirm key params** (skip what's already clear): duration (4–8 / 9–12 / 13–15 /
   >15 → segmented); aspect (16:9 / 9:16 / auto); assets (text-only / images / images +
   clips); mood, camera style, use case.
3. **Write the prompt** — ≤15s: offer 2–3 distinct styled versions; >15s: a segmented
   plan. Every prompt is copy-ready and maps assets to API parameters.
4. **Render** — once the user picks a version (and provides any asset URLs), call the
   API via `scripts/generate.sh` and return the `video_url`.
5. **Iterate** — adjust a beat, swap style/palette/camera, add/cut lines/SFX, change
   duration or segmentation, then re-render.

## Notes

- Write vivid, concrete, natural language — Seedance 3.0 parses it well.
- Always describe each asset's role and map it to `image_url` / `end_image_url` /
  `media_urls`; don't mix up which image is which.
- Be explicit whether you're *referencing* (borrow style/motion) or *editing*
  (modify the source clip).
- Match image-gen style to the theme: wuxia → 3D Chinese-fantasy render; historical →
  ink wash / gongbi; sci-fi/cyberpunk → futuristic realistic CG; people/realism →
  cinematic photography; food → commercial food photography; nature → landscape /
  aerial documentary.
- Order actions and camera moves in time so the model understands sequence.
- For 13–15s, prefer ordered shot breakdown; use timestamps only when explicitly useful.
- Quote dialogue and label speaker + emotion; put SFX on its own line.
- Keep prompts focused — emphasize the core, avoid information overload. Mood and
  atmosphere strongly affect the result; don't skip them.