Seedance 2.0 Superprompt
scotti1i/seedance-2.0-superprompt
Escreva, audite e corrija prompts Seedance 2.0 com regras estruturadas
Ver no GitHubSkill do Claude Code para escrever, auditar e corrigir prompts de geração de vídeo para ByteDance Seedance 2.0. Oferece três modos (WRITE/LINT/FIX), 8 templates de produção e 25 regras anti-padrões que transformam ideias brutas em prompts prontos para uso. Sem dependências de CLI, API ou credenciais.
Instalação
npx convez-ai scotti1i/seedance-2.0-superprompt1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar prompts estruturados para vídeos em Seedance 2.0
- Auditar e pontuar prompts existentes contra 25 regras de qualidade
- Reparar prompts falhos com explicação de mudanças
- Consultar templates de produção para diferentes cenários
- Aplicar metodologia de prompt portável em qualquer LLM
Veredito

Ferramenta especializada e bem estruturada para criadores de vídeo em Seedance 2.0. A abordagem baseada em regras e workflows (WRITE/LINT/FIX) é mais útil que galerias de exemplos para produção repetível, mas o escopo é bastante nicho—apenas relevante para usuários de ByteDance.
README
A tradução para o português deste README ainda está na fila.
Seedance 2.0 Superprompt
A prompt operating system for ByteDance Seedance 2.0: write new video prompts, audit weak prompts, and fix broken prompts into paste-ready production prompts.
Unlike prompt galleries, this repo does not try to collect thousands of examples. It gives you the reusable workflow behind good Seedance prompts: rules, templates, decision trees, and a Claude Code skill that applies them for you.
Turn any rough video idea — even a junk-token soup — into a production-grade Seedance 2.0 prompt that pastes anywhere. Zero CLI · Zero API · Zero credentials.
⭐ Star on GitHub · 📚 Methodology · 🧩 Template catalog · 🇨🇳 中文
Why this repo exists
Prompt galleries are useful when you need inspiration. This repo is for the next step: turning a messy idea into a repeatable, reviewable prompt workflow.
| If you want... | Use this |
|---|---|
| A prompt for a new idea | WRITE mode generates a structured 3-segment prompt |
| To know why a prompt fails | LINT mode scores it against 25 rules |
| To repair a bad prompt | FIX mode rewrites it and explains the diff |
| A proven starting point | 8 templates cover product, pet, food, city walk, unboxing, and more |
| A portable rulebook | Use methodology.md, lint-rules.md, and templates in any LLM |
Status: solo-maintained by @scotti1i. Free to use, fork, and redistribute (MIT). PRs not accepted at this stage — please use Issues for bugs, lint-rule proposals, and template ideas. See CONTRIBUTING.md.
⚡ Install in 30 seconds
git clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cd seedance-2.0-superprompt && bash install.sh
Restart Claude Code. Type /seedance and you're live.
How people use this repo
| User type | Best path |
|---|---|
| Just want to use the skill | Run the install command above |
| Want to inspect the rules | Read methodology.md and lint-rules.md |
| Want reusable examples | Start from the template catalog |
| Want to customize it | Fork this repo to your own account and edit the seedance/ folder |
| Repo owner | You cannot fork your own repo into the same account; clone it locally instead |
🛠️ Works with
This is portable prompt knowledge, not Claude Code-locked. The
SKILL.md + methodology.md + templates/ + lint-rules.md files
work as a reference library inside any AI coding agent.
| Tool | How to use |
|---|---|
| Claude Code (primary) | Auto-install via bash install.sh → use /seedance ... natively |
| Codex CLI | Paste seedance/SKILL.md content as the agent's system prompt; reference seedance/methodology.md for rules |
| Cursor | Drop the seedance/ folder into your project; ask Cursor to read it and follow the workflow in SKILL.md |
| Claude.ai / Claude.app | Upload methodology.md + relevant template as project knowledge |
| ChatGPT / Gemini / DeepSeek | Copy SKILL.md + methodology.md into the chat; ask the model to follow them |
| Manual | Read templates/INDEX.md, copy a template, fill variables yourself, paste into your Seedance client |
The actual output is always plain text prompts. The skill mechanic is the workflow; the methodology and templates are portable knowledge.
🎯 What you get (3 modes, 1 skill)
| You type | Skill returns |
|---|---|
/seedance 10s vertical cat vlog in afternoon light | A full 3-segment engineered prompt + parameters + paste-target list |
/seedance lint <your prompt> | A rule-by-rule audit + 1-10 score in 5 seconds |
/seedance fix <your bad prompt> | Lint report + auto-rewritten prompt + diff explanation (the killer feature) |
Bottom line: stop writing AI-generic prompts. This skill enforces
named cameras (Sony A7S3, ARRI Alexa, iPhone 15 Pro), 3-segment
timelines, color+material+light scene triplets, @image rebinding, and
3-5 specific negatives — automatically, every time, citing the official
ByteDance methodology rule for each decision.
The output is plain text. Paste it into: 即梦 · CapCut Dreamina · Volcano Engine API · BytePlus ModelArk · PiAPI · Coze · or any third-party wrapper.
⭐ Headline demo: FIX a junk prompt in 5 seconds
You type (typical newbie prompt)
A beautiful young woman drinking coffee in a cozy cafe, cinematic,
highly detailed, masterpiece, 8k, amazing quality
Skill responds
## Lint report — score: 1/10
🔴 R-J01: junk tokens "cinematic" / "highly detailed" / "masterpiece" / "8k"
🔴 R-S01: missing color + material + light triplet
🔴 R-T01: no time segments → pacing will be random
🔴 R-A01: no Overall audio atmosphere → BGM ignored
🔴 R-C02: no camera move
🔴 R-J03: no named camera
🔴 R-N01: no negatives
... (9 violations total)
## Fixed prompt (paste this)
Vertical video 9:16. Overall audio atmosphere: warm lo-fi piano with
soft espresso machine hiss, distant cafe chatter, no voiceover.
[0-3s] Medium close-up handheld shot. A young east-asian woman with
shoulder-length dark hair, cream cashmere sweater, sits at a marble
counter. Warm tungsten pendant light overhead, dark walnut shelves with
mason jars behind, single ceramic mug with steam rising. Slow push-in
toward the mug. Sound: porcelain settles on marble, espresso machine
hisses softly.
[3-7s] Cut to extreme close-up of the mug. Her fingers wrap around the
ceramic, condensation traces the rim, golden window light catches the
foam ripples. She tilts the mug slowly. Sound: gentle slosh, lo-fi piano
swells.
[7-10s] Pull back to medium shot. She lifts the mug, takes a slow sip,
eyes drift to the window, corner of her mouth lifts into a small private
smile. Sound: piano resolves into a soft chime.
Shot on Sony A7S3, natural skin tone, slight film grain, documentary
realism. Face stable, no distortion, motion fluid, no plastic skin,
no oversaturated colors, no on-screen text.
## What changed and why
Removed `cinematic` / `8k` / `masterpiece` → added `Shot on Sony A7S3`
(R-J01 / R-J03 — junk tokens hurt, named camera invokes documentary realism)
Added 3-segment timeline `[0-3s] [3-7s] [7-10s]` (R-T01)
Added color + material + light triplet to every scene (R-S01)
Added one camera move per segment: push-in → locked → pull back (R-C01)
Added `Overall audio atmosphere:` as line 2 (R-A01)
Added 5 specific negatives at the end (R-N01)
... (full diff in skill output)
## Mode & parameters
- Mode: T2V (no assets implied)
- Duration: 10s · Ratio: 9:16 · Model: seedance2.0fast
- Paste into: jimeng.jianying.com, dreamina.capcut.com, Volcengine API,
BytePlus, PiAPI, Coze, etc.
Score: 1/10 → 10/10. ~30 seconds.
Now paste the fixed prompt into jimeng.jianying.com, dreamina.capcut.com, Volcengine API, BytePlus, or any wrapper. Done.
See the full case study.
Three workflows
1. WRITE — idea → prompt (default)
/seedance 10s vertical TikTok-style coffee shop POV in autumn
→ Skill picks T2V mode, writes 3-segment engineered prompt, lists clients.
2. LINT — audit an existing prompt
/seedance lint <your prompt>
→ Skill returns rule-by-rule report + 1-10 score. No rewriting.
3. FIX — rewrite a bad prompt (the killer feature)
/seedance fix <your bad prompt>
→ Lint report + REWRITTEN prompt + diff explanation (what changed, why).
All three workflows output text only — paste into your Seedance client of choice.
Methodology behind the skill
Seedance 2.0 by ByteDance is the strongest open multimodal video model available — but writing good prompts for it is a different skill from writing prompts for SD / Midjourney / Veo. The official guide is ~1100 lines of methodology. The reward for reading it is real (10× better output) but the entry barrier is high.
This skill compresses that 1100-line SOT into a Claude-driven workflow:
| Problem | Fix |
|---|---|
Junk tokens like highly detailed, 8k, masterpiece hurt output | Lint catches them, replaces with named cameras (Sony A7S3, ARRI Alexa, iPhone 15 Pro, 35mm film) |
| Too many time segments (≥5) silently collapse | Skill forces 3-segment structure |
| I2V users re-describe the subject and cause identity drift | Skill enforces "motion only, never subject" |
R2V users bind @image1 once and get face swaps mid-shot | Skill rebinds in every segment |
Multi-character prompts use woman 3 and parser breaks | Skill uses letters A/B/C |
| Vague action verbs ("interacts with the product") produce vague output | Skill rewrites to concrete sequences ("rotates the bottle, places it on the table") |
Install
Option A — One-line install
curl -fsSL https://raw.githubusercontent.com/scotti1i/seedance-2.0-superprompt/main/install.sh | bash
Option B — Manual
git clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cp -r seedance-2.0-superprompt/seedance ~/.claude/skills/seedance
Option C — Symlink (for contributors)
git clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cd seedance-2.0-superprompt
ln -s "$(pwd)/seedance" ~/.claude/skills/seedance
After install, restart Claude Code. The skill becomes available as
/seedance and is auto-triggered when you mention Seedance / video
generation / Dreamina / 即梦.
Quick start
🪄 Fix a bad prompt (most popular)
/seedance fix a beautiful girl drinking coffee, cinematic, 8k, masterpiece
→ Returns lint report + full rewrite + diff explanation in one pass. See the headline demo.
Speed mode (one sentence)
/seedance 10s vertical cat playing with a ball in soft afternoon light
→ Skill auto-picks T2V, 9:16, 10s, seedance2.0fast, generates a 3-segment engineered prompt, and lists where you can paste it.
Precision mode (lock specific knobs)
/seedance product showcase 8s 16:9 cinematic, ARRI Alexa look, no subtitles
→ Skill respects every constraint you specify.
Templates (pre-built scenarios)
/seedance use template pet-vlog, my pet is a tabby cat, product is a blue
collar from Petsmart
→ Substitutes variables in the verified pet-vlog template.
Lint only (don't rewrite)
/seedance lint <your prompt>
→ Runs 25 lint rules and reports a grade. No rewriting.
Features
🪄 Three workflows in one skill
- WRITE: idea → engineered prompt
- LINT: audit any prompt against 25 anti-pattern rules
- FIX: rewrite a bad prompt into a good one, with diff explanation
🎯 Auto mode selection
Picks T2V / I2V / R2V / V2V / Frames based on what assets you have.
📐 Methodology enforcement
25 lint rules distilled from the official ByteDance / Volcano Engine prompting guides and community-published field reports.
📚 Template library
8 verified production templates: pet vlog, product showcase, morning routine, city walk, food ASMR, ancient drama, unboxing, before/after.
🌍 Multi-client output
One prompt format works across:
- jimeng.jianying.com (China web)
- dreamina.capcut.com (Global web)
- Volcano Engine API (
doubao-seedance-1-0-pro-250528) - BytePlus ModelArk
- PiAPI, Coze, 即梦助手 (third-party wrappers)
- ByteDance internal CLI (if you have access)
🚫 Zero infrastructure
No API calls. No credentials. No payment. Pure prompt engineering.
🌐 Bilingual
Prompts in Chinese, English, or mixed. Engineering tags stay English (industry standard), content language is your choice.
Repository structure
seedance-2.0-superprompt/
├── README.md ← you are here
├── LICENSE ← MIT
├── CONTRIBUTING.md ← how to add templates / lint rules
├── install.sh ← one-line installer
├── TESTING.md ← end-to-end test scenarios
├── .github/ ← issue and PR templates
└── seedance/ ← the skill itself (symlink target)
├── SKILL.md ← Claude entry point
├── methodology.md ← 500-line distilled SOT
├── modes/ ← per-mode handbooks (t2v / i2v / r2v / v2v / frames)
├── decision-trees/ ← visual pickers (mode / duration / camera / safety)
├── templates/ ← verified production templates
├── examples/ ← case studies and prompt examples
└── lint-rules.md ← 25 anti-pattern catalog
Template catalog
| Template | Mode | Duration | Use case |
|---|---|---|---|
| pet-vlog | R2V | 15s | Pet outdoor walk + product showcase (TikTok viral) |
| product-showcase | R2V | 10s | Hero rotation + features + brand close-up |
| morning-routine | T2V | 10s | Slow-cinematic lifestyle vignette |
| city-walk | T2V | 10s | POV travel through aesthetic district |
| food-asmr | T2V | 8s | Food close-up with ASMR sound |
| ancient-drama | R2V | 15s | Wuxia two-character confrontation |
| unboxing | R2V | 10s | POV unboxing reveal |
| before-after | R2V | 8s | Cosmetics / transformation reveal |
Add yours with a PR — see CONTRIBUTING.md.
Methodology summary
The skill enforces:
- Eight-element formula — subject + action + scene + light + camera + style + quality + constraints, in that order
- Three-segment timeline —
[0-Xs][X-Ys][Y-Zs]. ≥5 collapses. - Scene triplet — color + material + light for every scene
- One camera move per segment — no stacking
- Named camera, not "cinematic" — Sony A7S3, ARRI Alexa, iPhone 15 Pro, 35mm film
- R2V:
@image_Nrebound every segment — pronouns lose anchor - Multi-character: letters A/B/C — numbers confuse the parser
- I2V: never re-describe subject — image already has the info
- Audio atmosphere declared at top — buried BGM gets ignored
- 3-5 specific negatives — more over-constrains
Full details: methodology.md.
What this skill is NOT
- ❌ A wrapper around the dreamina / volcengine CLI
- ❌ A video generator (it produces prompts, not videos)
- ❌ A paid service (no credits, no API keys)
- ❌ A guarantee — Seedance is non-deterministic; expect 2-5 retries for hero shots
- ❌ Affiliated with ByteDance / Volcano Engine / Dreamina / Jianying
👤 About the author
ScottLi (@scotti1i) — AI builder working on TikTok Shop, cross-border e-commerce, and AI video tooling. I open-source the methodology tools I build for my own work.
If this skill helps you, give it a ⭐ on GitHub and reach out via any channel below — I read everything.
| Platform | Handle |
|---|---|
| yifanlitju@gmail.com | |
| ▶️ YouTube | @ScottGlobalAI |
| 🎵 抖音 (Douyin) | SCT出海Scott |
| 🐙 GitHub | @scotti1i |
Sponsor / hire
If your team is building with Seedance 2.0 or other ByteDance video models at scale and wants a custom prompt-engineering audit, methodology training, or workflow integration — get in touch via any of the channels above.
Contributing
The skill gets stronger every time someone:
- adds a new verified template
- finds a new anti-pattern and writes a lint rule
- updates methodology after a Seedance model version bump
- contributes a good / bad example
See CONTRIBUTING.md and the issue templates to get started.
Roadmap
- Add 12 more templates (target: 20 total)
- Cross-model adapters (when Veo / Sora prompts diverge)
- Auto-update when official ByteDance SOT changes
- Video-output benchmark suite (with anonymized eval samples)
- CLI companion (lint as a standalone binary)
License
MIT. See LICENSE.
Acknowledgments
- ByteDance Seedance team for the official prompting guide
- Volcano Engine docs team for the API reference
- The community of creators who've shared field-tested prompts on Lark, WeChat, and small Discord servers — your write-ups made this possible