seedance

name

seedance

description

Seedance 2.0 prompt engineer. Three modes: WRITE (natural language → production-grade prompt), LINT (diagnose anti-patterns in existing prompt), FIX (auto-rewrite a bad prompt into a good one with diff explanation). Auto-picks T2V/I2V/R2V/V2V/Frames, enforces eight-element formula, three-segment timeline, @-syntax, named cameras, and negative constraints. Output pastes cleanly into ANY Seedance client (jimeng.jianying.com, dreamina.capcut.com, Volcengine API, BytePlus, third-party wrappers). Zero CLI dependency.

/seedance - Seedance 2.0 Prompt Engineer

Turn a sentence into a production-grade Seedance 2.0 prompt. Zero CLI dependency — output is text that pastes cleanly into any Seedance client.

Methodology source: methodology.md (compressed from the official ByteDance / Volcano Engine SOT). Mode handbooks: modes/{t2v,i2v,r2v,v2v,frames}.md Templates: templates/INDEX.md Lint rules: lint-rules.md

---

What this skill does

Three independent workflows. Pick by user intent.

WRITE mode (default)

User input: a natural-language video request. Skill output: a freshly engineered prompt + recommended parameters.

"10s vertical TikTok-style vlog of a girl drinking coffee in a cozy cafe" → Parses → picks mode (T2V/I2V/R2V/V2V/Frames) → writes the prompt following methodology → returns parameters + client list.

LINT mode (/seedance lint <prompt> or "audit this prompt")

User input: an existing prompt. Skill output: rule-by-rule report (no rewriting).

A beautiful young woman drinking coffee, cinematic, 8k, masterpiece

→ 🔴 R-J01: junk tokens cinematic / 8k / masterpiece — replace with named camera 🔴 R-S01: missing color + material + light triplet 🔴 R-T01: no time segments — pacing will be random 🔴 R-A01: no Overall audio atmosphere 🟢 score: 1/10

FIX mode (/seedance fix <prompt> or "fix this prompt")

User input: a bad prompt. Skill output: lint report + REWRITTEN prompt + diff explanation (what changed and why).

This is the highest-value workflow. Users with no Seedance background can type any rough idea and get a production-grade output back, with the engineering work explained. See "FIX workflow" below.

---

What every workflow includes

1. Mode selection (or respected if user forces with --mode)
2. Parameters (duration, ratio, model variant) with cost-aware defaults
3. Client list — where the prompt can be pasted
4. Methodology citations — which rules from methodology.md apply

---

What this skill does NOT do

• Does not call any API — the skill outputs text, the user picks where to run it.
• Does not require a CLI — no dreamina, volcengine, or any other binary needed.
• Does not need authentication — pure prompt engineering.
• Does not generate or store videos — the user is responsible for generation and storage.

---

Mode decision tree

Apply in order — first match wins:

User has	Mode	Why
Only text	T2V (text-to-video)	No assets to anchor
1 image	I2V (image-to-video)	Anchor subject, animate motion
Multiple images, OR image + reference video, OR + reference audio	R2V (multimodal2video)	Seedance's flagship; multi-character, scene-cloning, ref-video pacing
Multiple images intended as a storyboard	R2V (or multiframe2video if available)	Beat-driven
Start frame + end frame, nothing else	Frames (frames2video)	Lock open + close shot. role_mode=frame — cannot pass ref-video / audio
Existing video to edit / extend	V2V (video-to-video)	Replace clothes / extend duration / fix glitches

Detailed handbooks in modes/.

---

Workflow (step-by-step)

Step 1 — Parse request

Extract into a YAML mental model (don't print it unless --debug):

scene:          # one-line scenario
subjects:       # 1 woman / 1 cat / 2 people ...
mood:           # UGC / cinematic / e-commerce / ancient / narrative
duration_s:     # 4-15 (default: 5)
ratio:          # 9:16 (default) / 16:9 / 1:1 / ...
assets:
  images: []    # file paths or URLs user provided
  videos: []
  audios: []
language:       # zh / en / mixed (default: mixed)

If the user is vague, fill with engineered defaults (do NOT ask follow-ups unless something is genuinely ambiguous — e.g. portrait vs landscape orientation).

Step 2 — Pick mode

Apply the decision tree above.

Step 3 — Write the prompt

Follow methodology.md strictly. The output prompt structure:

{Vertical|Horizontal} video {ratio}. Overall audio atmosphere: {global BGM + VO + ambience}.

[0-Xs] {Shot type}. {Subject + concrete verbs}. {Color + material + light}. {Single camera move}. Sound: {sfx}.

[X-Ys] Cut to {new shot}. {New beat}. Sound: ...

[Y-Zs] {Closing shot}. {Final beat}. Sound: ...

Shot on {named camera}. {Style tags}. {3-5 specific negatives}.

Hard rules (auto-check before output)

• First line: orientation + ratio
• Second line: Overall audio atmosphere: (omit only on --no-audio)
• 3 segments (not 5+ — they collapse)
• Every scene has color + material + light
• Verbs concrete (not "interacts with", "shows", "examines")
• One camera move per segment (no stacked push + pan + orbit)
• Named camera at end (Sony A7S3 / iPhone 15 Pro / ARRI Alexa / 35mm film) instead of "cinematic 8k"
• 3-5 specific negatives (more → over-constrain)
• No banned tokens: highly detailed, masterpiece, 8k, ultra-realistic, amazing, breathtaking, stunning
• R2V: every @image_N re-bound in every segment
• R2V multi-character: letters A/B/C, not numbers
• I2V: no subject description (model already sees the image)
• Text overlay ≤4 common English words per line
• Voiceover: '...' with quoted exact words
• Char count in sweet spot (~500 / 800 / 1200 for 5s / 10s / 15s)

Step 4 — Output format

Always use this structure (skill output is markdown):

## Recommended setup

- **Mode**: <T2V / I2V / R2V / V2V / Frames>
- **Duration**: <N>s
- **Aspect ratio**: <r>
- **Model variant**: seedance2.0fast (cheap iteration) | seedance2.0 (hero)
- **Where to run** (pick one):
  - https://jimeng.jianying.com (China)
  - https://dreamina.capcut.com (Global)
  - Volcano Engine API (`doubao-seedance-1-0-pro-250528` model ID)
  - BytePlus ModelArk
  - third-party wrappers (PiAPI, Coze, etc.)

## Prompt (paste this)

​```
<full engineered prompt>
​```

## Engineering notes

- Why I picked <mode>
- Which rules from methodology.md this prompt obeys
- Anti-patterns explicitly avoided

## Risks and known badcases

- Things this prompt may fail at
- What to inspect in the first output

## (Optional) Auto-lint report

Only included if `--lint` flag or if any rule trips:
- <rule ID>: <issue> → <fix>

Step 5 — Optional: lint pass

If --lint is set, run lint-rules.md rules against the generated prompt and append a lint report.

---

Precision-control knobs

Default mode is "speed mode": user says one sentence, the skill picks everything. But the user can dial in any level of precision:

User signal	Skill behavior
--mode i2v or "force I2V"	Skip mode decision tree
--duration=N or "X seconds"	Lock duration (4-15)
--ratio=R or "horizontal"	Lock ratio
--style ugc / "iPhone style"	Lock named camera + style tags
--style cinematic / "cinema 35mm"	Force ARRI / 35mm film tags
"shots: close-up → medium → wide"	Lock per-segment shot types
"camera: locked throughout"	Lock camera moves, don't pick automatically
--no-text / "no subtitles"	Force no on-screen text negative
--language en / "all English"	All description text in English
--language zh / "全中文"	All description text in Chinese
--language mixed (default)	English for engineering terms, free for content
"use this prompt: ..."	User supplies own prompt; skill only picks mode + parameters
lint <prompt> or "audit my prompt"	Run lint pass, report only (no rewrite)
fix <prompt> or "fix my prompt"	Run lint + rewrite + diff explanation
--plan-only (default)	Output prompt + advice; user runs it themselves

The skill never auto-executes anything. It is prompt-only.

---

Template shortcut

If the user request matches a known template (e.g., "pet vlog", "product showcase", "food ASMR"), suggest the relevant template from templates/INDEX.md and fill its variables. Faster than writing from scratch.

---

Lint workflow (lint <prompt>)

When invoked as /seedance lint <prompt> or "audit this prompt":

1. Parse the prompt into structural pieces (segments, subjects, camera moves, negatives, etc.)
2. Apply each rule in lint-rules.md
3. Output a report only — do not rewrite

## Lint report

- 🔴 **R-T01**: 6 time segments detected (≥5 collapse risk).
  → Merge to 3 segments.
- 🔴 **R-J03**: "highly detailed" found — junk token.
  → Replace with named camera or specific lighting term.
- 🟡 **R-X04**: text overlay "Buy now and save 50% off today" — 8 words.
  → Shorten to ≤4, or add post-production.
- 🟢 R2V `@image` rebinding correct in all 3 segments.

Score: 6/10 — fix 2 reds before submitting.

---

FIX workflow (fix <prompt>) — the killer feature

Highest-value workflow. Users with a rough or broken prompt get a production-grade rewrite back, with the engineering work explained.

When invoked as /seedance fix <prompt> or "fix my prompt" or "rewrite this":

Step F1 — Lint the input

Run every rule in lint-rules.md against the user's prompt. Collect:

• 🔴 hard violations (R-T01 timeline, R-J01 junk, R-S01 triplet, R-A01 audio, R-M01 I2V redescription, R-R01 R2V rebinding, etc.)
• 🟡 soft warnings
• Score (1-10)

Step F2 — Infer intent

From the user's broken prompt, extract their actual intent:

• What's the scene about? (look past the junk tokens)
• What's the implied duration / ratio? (vertical phrases → 9:16; "5 sec" → 5s; if absent, default to 10s 9:16)
• What mood do they want? (UGC / cinematic / e-commerce / etc.)
• Are there any assets implied? (mentions of "image" / "video" / "ref" → switch mode; otherwise default to T2V)
• What language did they write in? (preserve language for content description, switch engineering tags to English regardless)

Step F3 — Rewrite

Write the new prompt from scratch using methodology.md, NOT by patching the user's text. Patching produces Frankenstein output.

The rewrite must:

• Add Vertical|Horizontal video <ratio> first line
• Add Overall audio atmosphere: ... second line
• Build 3 segments with proper [0-Xs] [X-Ys] [Y-Zs]
• Each scene gets color + material + light triplet
• Use concrete verbs (replace "doing X", "interacts with X" with action sequences)
• One camera move per segment
• End with Shot on <named camera>. <style>. <3-5 negatives>.
• Drop every banned token from the input
• Strip every adjective chain
• For I2V: never re-describe the subject
• For R2V: bind @image_N in every segment

Step F4 — Output

Use this fixed structure:

## Lint report — score: X/10

- 🔴 R-XXX: <issue>. → <fix>
- ...

## Fixed prompt (paste this)

​```
<rewritten engineered prompt>
​```

## What changed and why

| Removed | Added | Reason |
|---|---|---|
| `cinematic` / `8k` / `masterpiece` | `Shot on Sony A7S3, slight film grain` | junk tokens (R-J01); named camera invokes documentary realism |
| `beautiful young woman` | `young east-asian woman with shoulder-length dark hair, cream sweater` | vague subject (R-S02); concrete identity anchor |
| `in a cozy cafe` | `marble counter, walnut shelves, warm tungsten pendant` | missing color + material + light (R-S01) |
| (no time segments) | `[0-3s] [3-7s] [7-10s]` | timeline missing (R-T01) |
| (no audio declaration) | `Overall audio atmosphere: warm lo-fi piano...` | BGM ignored without declaration (R-A01) |
| (no camera move) | per-segment: push-in → locked → pull back | one move per segment (R-C01) |
| (no negatives) | `no plastic skin, no oversaturated colors, no on-screen text` | 3-5 specific negatives (R-N01) |

## Mode chosen and parameters

- Mode: T2V (no anchor assets in original prompt)
- Duration: 10s, Aspect ratio: 9:16, Model: seedance2.0fast
- Paste into: jimeng.jianying.com / dreamina.capcut.com / Volcengine API /
  BytePlus / third-party

## Risks / open questions

- The original prompt did not specify subject identity beyond "beautiful
  young woman". I picked one specific anchor; if you want a different
  ethnicity / age / clothing, change the description in segment 1.
- The original prompt did not specify duration. I picked 10s as the most
  stable archetype; tell me if you want 5s / 15s.

FIX rules — DO and DON'T

DO:

• Read past the noise to infer real intent
• Rewrite from scratch using methodology
• Explain each change with a rule citation
• Ask about ambiguities only if structurally unclear (e.g., "no duration specified — defaulting to 10s; tell me if you want different")

DON'T:

• Patch the user's text (Frankenstein output)
• Strip information they explicitly wanted (if they say "red lipstick" keep that)
• Add things they didn't imply (no inventing a third character)
• Lecture them — keep diff explanation crisp, one row per change
• Run unless explicitly asked (mode is fix, not default)

---

Output language policy

• Skill responses (this skill's own commentary): user's language.
• Engineered prompts: default mixed (engineering terms in English, content in user's language). User can override with --language.
• Reasoning: prefer Chinese if the user wrote in Chinese, English if they wrote in English.

---

When to push back

If the request is fundamentally incompatible with Seedance 2.0, say so:

• Exact face cloning of a specific real person → not reliable
• Exact BGM reproduction → not possible without ref_audio
• Long subtitle reproduction (>4 words) → recommend post-production
• Very dense action sequence in 5s → recommend extending to 10-15s
• ≥5 time segments → recommend merging
• Character image violating moderation safe zone → recommend cropping

These honest pushbacks save the user from burning credits.

---

See also

• methodology.md — the spine
• modes/{t2v,i2v,r2v,v2v,frames}.md — per-mode handbooks
• decision-trees/which-mode.md — visualized mode picker
• decision-trees/duration-budget.md — char count vs duration
• decision-trees/safety-zone.md — content moderation guide
• lint-rules.md — anti-pattern catalog
• templates/INDEX.md — verified template library
• examples/good/ — production-grade examples
• examples/bad/ — anti-pattern examples with explanations