English · 中文 · Methodology · Templates · YouTube
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
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.
git clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cd seedance-2.0-superprompt && bash install.shRestart Claude Code. Type /seedance and you're live.
| 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 |
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.
| 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.
A beautiful young woman drinking coffee in a cozy cafe, cinematic,
highly detailed, masterpiece, 8k, amazing quality
## 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.
/seedance 10s vertical TikTok-style coffee shop POV in autumn
→ Skill picks T2V mode, writes 3-segment engineered prompt, lists clients.
/seedance lint <your prompt>
→ Skill returns rule-by-rule report + 1-10 score. No rewriting.
/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.
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") |
curl -fsSL https://raw.githubusercontent.com/scotti1i/seedance-2.0-superprompt/main/install.sh | bashgit clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cp -r seedance-2.0-superprompt/seedance ~/.claude/skills/seedancegit clone https://github.com/scotti1i/seedance-2.0-superprompt.git
cd seedance-2.0-superprompt
ln -s "$(pwd)/seedance" ~/.claude/skills/seedanceAfter install, restart Claude Code. The skill becomes available as
/seedance and is auto-triggered when you mention Seedance / video
generation / Dreamina / 即梦.
/seedance fix a beautiful girl drinking coffee, cinematic, 8k, masterpiece
→ Returns lint report + full rewrite + diff explanation in one pass. See the headline demo.
/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.
/seedance product showcase 8s 16:9 cinematic, ARRI Alexa look, no subtitles
→ Skill respects every constraint you specify.
/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.
/seedance lint <your prompt>
→ Runs 25 lint rules and reports a grade. No rewriting.
- 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
Picks T2V / I2V / R2V / V2V / Frames based on what assets you have.
25 lint rules distilled from the official ByteDance / Volcano Engine prompting guides and community-published field reports.
8 verified production templates: pet vlog, product showcase, morning routine, city walk, food ASMR, ancient drama, unboxing, before/after.
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)
No API calls. No credentials. No payment. Pure prompt engineering.
Prompts in Chinese, English, or mixed. Engineering tags stay English (industry standard), content language is your choice.
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 | 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.
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.
- ❌ 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
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 | |
| @ScottGlobalAI | |
| 🎵 抖音 (Douyin) | SCT出海Scott |
| 🐙 GitHub | @scotti1i |
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.
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.
- 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)
MIT. See LICENSE.
- 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