English | 中文 | 日本語 | 한국어 | Español | Português | Français

Cinematic Shot Language Designer

Turn your vague video ideas into cinema-grade video prompts ready for Jimeng Seedance 2.0 — in one shot.

A Claude Skill built on the Agent Skills specification, blending Hollywood-level cinematography aesthetics with Chinese film industry practices. It's designed to help creators break free from the "looks nice but random" AI video trap and achieve precise, controllable visual storytelling.

# Clone into your project's Skill directory
git clone https://github.com/woodfantasy/Seedance2.0-ShotDesign-Skills.git .claude/skills/seedance-shot-design

Please learn this skill: https://github.com/woodfantasy/Seedance2.0-ShotDesign-Skills

git clone https://github.com/woodfantasy/Seedance2.0-ShotDesign-Skills.git agents/skills/seedance-shot-design

git clone https://github.com/woodfantasy/Seedance2.0-ShotDesign-Skills.git .cursor/skills/seedance-shot-design

Just tell Claude:

Write a 15-second cyberpunk rain chase video prompt

The Skill auto-activates and generates the prompt through a 5-step workflow:

1. Requirement Analysis — Confirm duration / aspect ratio / assets / style
2. Visual Diagnosis — Select camera language & director style
3. Six-Element Assembly — Compose the prompt using the structured formula
4. Mandatory Validation — Apply 7-rule structured validation checklist
5. Professional Delivery — Director's statement + complete prompt

User Input:

Write a 10-second Eastern Xianxia short film video prompt

Skill Output:

Seedance Video Prompt

Theme: A young man in white catches a red leaf in a misty ancient temple at dawn, reaching enlightenment

Director's Statement (for creative intent only — do not copy): Uses an aerial → dolly → slow push three-phase camera progression, transitioning from grand vista to intimate emotion. 35mm film grain adds a handcrafted texture, and the gold-teal palette echoes the Eastern philosophy of harmony with nature.

Full Prompt (copy directly into Jimeng's input field):

10-second Chinese fantasy, realistic Eastern cinematic look, gold-teal palette, ethereal ambient sound.
0-3s: High-angle aerial shot of an ancient temple amid clouds, slow aerial push, morning mist drifting through valleys, distant bell faintly ringing, Tyndall light beams piercing through cloud layers.
3-7s: Dolly push through the temple gate into the courtyard, a young man in white raises his hand to catch a falling red leaf, 35mm film grain texture, shallow depth of field focusing on hand details.
7-10s: Close-up of the young man looking up, slow push-in, wind rises, sleeves and hair sweep to the right side of frame, spiritual light spiraling upward in the courtyard.
Sound: Ambient sound contracts into a single clear sword ring.
Forbidden: Any text, subtitles, logos, or watermarks

# AI Comic Drama
Write a 10-second AI comic-style bossy CEO short, vertical 9:16, with dialogue and exaggerated close-up expressions

# E-commerce Ad
Write an 8-second luxury watch product ad video prompt, 9:16 vertical

# Short Drama Dialogue
Write a 12-second plot-twist short drama scene with dialogue

# One-Take
Write a 15-second one-take museum walkthrough video prompt

# Image-to-Video (I2V)
I have a character design image — generate a 10-second animation from this first frame

# Video Extension
Extend this video by 10 seconds with a sunset ending

# Multiframe Storytelling
I have 5 storyboard keyframe images — create a coherent story video that flows through all frames

# With Reference Assets
I've uploaded 3 character design images and 1 reference video — generate a 15-second Xianxia fight scene

seedance-shot-design/
├── SKILL.md                     # Core instructions (the Skill's brain)
├── README.md                    # This file
├── scripts/
│   ├── validate_prompt.py       # Industrial-grade prompt validation script
│   └── test_validate.py         # Validation script test cases
└── references/
    ├── cinematography.md        # Camera & focal length dictionary (incl. physical mounts & focal psychology)
    ├── director-styles.md       # Director style parameterized mapping (28+ styles, incl. Cel-Shaded CG)
    ├── seedance-specs.md        # Seedance 2.0 official platform specs
    ├── quality-anchors.md       # Quality anchors & lighting library (incl. NPR materials / lighting / conflict matrix)
    ├── scenarios.md             # Vertical scenario templates (20 scenarios + anime variants + video editing + multiframe storytelling + physics damping toolkit)
    └── audio-tags.md            # Audio & sound effect tag specs (incl. spatial acoustics & material-based onomatopoeia)

A standalone Python validation tool for developers and CI/CD pipelines. The AI agent applies these same rules natively via its built-in 7-rule validation checklist — no Python execution required during prompt generation.

# Validate text directly
python scripts/validate_prompt.py --text "your prompt"

# Validate from file
python scripts/validate_prompt.py --file prompt.txt

# Specify language (auto=auto-detect, cn=Chinese, en=English)
python scripts/validate_prompt.py --text "your prompt" --lang en

# JSON output (for programmatic processing)
python scripts/validate_prompt.py --text "your prompt" --json

Validation Checks:

• ❌ Over word limit (Chinese >500 chars / English >1000 words)
• ❌ Missing professional camera terminology
• ❌ Filler word hard-block (masterpiece / ultra-clear, etc. → error)
• ❌ Optical physics conflicts (ultra-wide + bokeh, handheld + perfect symmetry)
• ❌ Style conflict matrix (IMAX vs VHS, film vs digital, ink-wash vs UE5, Cel-Shaded vs realistic PBR, Slow Motion vs Speed Ramp)
• ❌ Asset reference overflow (images >9 / videos >3 / audio >3 / total >12)
• ❌ Long video (>5s) without time-slice hard-block
• ⚠️ Time-slice gaps or overlaps
• ⚠️ Declared duration vs slice endpoint mismatch
• ⚠️ In-segment motion logic conflicts
• ⚠️ Seedance review-risk bare English camera terms detection (Dolly / Aerial / Crane / Pan / Arc / Dutch / Steadicam)
• 🌐 Auto language detection (Chinese / English), adapting length standards & detection strategies per language
• 🎬 Multi-segment cross-segment consistency checks (style preamble / lighting structure / forbidden items)

Run Tests:

python -m unittest scripts.test_validate -v
# 54 tests pass (covering 11 test classes)

Following Agent Skills best practices:

• SKILL.md (~4000 tokens): Core workflow + structural templates + quality checklist
• references/ (three-layer routing): Camera dictionary and quality anchors are always loaded (Always-On); other knowledge bases auto-matched via semantic inference or loaded on explicit user request
• scripts/ (executed on demand): Validation runs only after prompt generation

• 🌐 Runway Platform Adaptation: Clarified native asset limits for Runway users (≤5 images, ≤3 videos) and provided explicit moderation bypass strategies for realistic human faces (blurring or NPR restyling).

• 🎞️ Start & End Frame Interpolation: Added a 7th multimodal reference pattern (@Image1 as start frame, @Image2 as end frame) to support precise point-to-point transitions.

• 🎬 Two New Effects Scenarios:

  • Freeze Time (Bullet Time): Camera dramatically weaves through completely frozen scene elements.
  • Multishot Video: Bypasses the "one-take" limitation by triggering the model to auto-generate sharp montage cuts in a single generation.

• 🚀 Extreme POV Scenarios: Added new 21st scenario template focusing on "human head-tracking logic", "high-speed projectile FPV (flying swords/arrows)", and "creature flight".

• 🎧 Immersive Audio Exclusion: Introduced strict audio-exclusion directives for POV templates (ONLY ambient sound, NO BGM or dialogue) to prevent AI from breaking immersion.

• 🧹 Background Purification Rule: Clarified that reference entity images must use a "pure white/blank background" to avoid contaminating the video environment in Image-to-Video generation.

• 🔗 CLI Integration Guide: New seedance-specs.md section mapping Shot Design modes to Jimeng CLI commands (text2video / image2video / multiframe2video / multimodal2video), with async task management and VIP channel documentation
• 🎞️ Multiframe Storytelling Template: New scenario template (#20) for multiframe2video — upload 2-9 keyframe images and let the engine auto-compose a coherent story video. Includes decision matrix for choosing between multiframe vs. multi-segment storyboard
• 📚 Knowledge Base Routing: Added multiframe and CLI routing entries to Step 2 semantic inference table

• 🎭 Descriptive Over Narrative Rule: New core rule (#12) — only write what the camera SEES (visual words), never what characters FEEL (emotion words). All emotions must be converted to visible physical expressions (facial micro-expressions, body language, breathing rhythm, gaze direction)
• ✍️ English Present Progressive: Assembly rules now mandate -ing form for English action descriptions (running not runs) — progressive tense implies continuous motion, matching video's dynamic nature
• 🎯 Motion Tone Front-Loading: Style preamble now explicitly declares overall motion energy (e.g., dynamic motion, high energy or serene, slow-paced atmosphere) to lock motion baseline early in generation

• 🎥 One-Shot-One-Move Rule: New core rule (#10) enforcing a single camera movement per time segment — combining movements (e.g., push-in + pan) causes jitter. Subject motion and camera motion must be described separately
• 🖼️ I2V Golden Rule: New core rule (#11) and dedicated I2V section — when generating video from an image, only describe motion/changes, never re-describe static content already in the first frame. Introduces preserve composition and colors anchor phrase
• 📏 Optimal Prompt Length: Added 60–100 word sweet spot guidance — below is vague, above causes concept drift and conflicting instructions
• 💪 Motion Intensity Modifiers: New bilingual quick-reference table in cinematography dictionary with 6 intensity tiers (violent → gentle → gradual) and do/don't examples to eliminate "mushy motion"
• 🎤 Rhythm Over Specs: Assembly rules now explicitly prefer semantic rhythm words (gentle/gradual/smooth) over technical parameters (24fps/f2.8) that Seedance cannot parse
• 🎬 Reference Video Best Practices: New practical constraints for reference clips — ideal 3–8s length, continuous shot (no cuts), single intent (subject OR camera, not both)

• 🛡️ Security Compliance: Resolved ClawHub OpenClaw "Suspicious patterns" flag by converting Python-based validation to LLM-native structured 7-rule validation checklist. Python scripts remain as standalone developer tools but are no longer executed by the agent
• 🎯 Trigger Phrase Optimization: Reduced activation trigger phrases from 40+ to 15 high-signal professional terms, lowering unintended activation surface while preserving core discoverability

• 🎤 Voice & Language Control System: New timbre cloning via video reference (语气和音色参考@视频1), dialect/accent control (Sichuan/Cantonese/Northeast/Taiwanese etc.), multilingual dialogue mixing, special voice styles (documentary narration / stand-up comedy / opera / ASMR)
• 📹 Multimodal Reference Guide: Upgraded from 4 brief tips to a structured guide with 6 core reference patterns (first frame / camera replication / action replication / camera+action separation / timbre reference / effect replication), plus multi-asset character control and consistency preservation guidance
• 📏 Video Extension Scenario: New forward/backward extension templates, seamless continuation techniques, duration cognition correction (generation duration = added seconds, not total)
• 📋 Story Completion Scenario: New storyboard-to-video, comic panel animation, and image-to-emotion-video creative modes
• 🎬 Creative Effects Quick Reference: New VFX trigger keywords — Hitchcock zoom, fisheye lens, particle effects, speed ramp, freeze transition, ink wash style, morphing effects (bilingual)
• 🎭 Emotion Performance Guidance: New emotion specificity table, emotion transition trigger words, emotion reference video usage in short drama chapter

• 🎯 Trigger Word Expansion: Massively expanded Skill activation coverage — added 20+ colloquial Chinese triggers (帮我写个视频, 拍一个, 做分镜, 短视频, AI视频, 抖音视频, vlog, 运镜...) and 10+ English triggers (generate a video, make a clip, shoot a scene, video script, drone shot, camera movement...) so the Skill auto-activates on natural, everyday user expressions — not just professional terminology

• 🔒 Security Compliance Optimization: Resolved ClawHub security flagging issues for shell execution, process control, and file access patterns while maintaining full functionality

• 🚨 Step 3 Mandatory Assembly Rules: Three-layer lighting must be on its own line with all three layers complete; SFX line must start with SFX:; prohibition line standardized (no custom additions); freestyle non-template sections forbidden
• ⛔ Step 4 Validation Blocking: Prompts failing validation are now forbidden from being shown to users; clear 5-step validation flow
• 📋 Step 5 Format Enforcement: Output must follow template exactly (Theme + Director's Note + code-block-wrapped prompt); missing any section = non-compliant
• 🎯 Step 2 Parameter Extraction Directive: Knowledge bases must not just be "loaded" — specific parameters must be extracted and embedded into the prompt

• 🧠 Smart Semantic Intent Routing: Step 2 knowledge base loading upgraded from "explicit trigger" to three-layer routing — Always-On loads camera dictionary & quality anchors every time, Semantic Intent Inference auto-detects needed knowledge bases from user's natural language, Explicit Override preserves direct user specification
• 🎯 Step 1 Smart Inference Principle: Agent proactively infers parameters (duration / style / scene) from a single user sentence, only asking about genuinely unknown info, limiting follow-up questions to 1-2
• 📝 Design philosophy updated from "loaded on demand" to "three-layer routing" ensuring every prompt has a quality foundation

• 🎭 Actor Blocking System: Three-element positioning (placement + face direction + gaze focus) with emotion modifier vocabulary for multi-character scenes
• 🎙️ Voiceover / Dialogue Split: Distinct templates for on-screen dialogue vs. off-screen voiceover / inner monologue, with anti-lip-sync directive for VO scenes
• 📐 Camera Angle Specificity: Vague → specific angle mapping (e.g., "close-up" → "over-shoulder medium close-up, focus on listener") with 5 comparison pairs
• 🎬 Narrative-Motivated Camera Movement: Camera moves now paired with storytelling purpose (e.g., "slow push-in — revealing inner turmoil")
• 🔀 Segment Transition Strategy: 6 transition types (gaze continuity / emotional escalation / contrast cut / spatial leap / temporal ellipsis / sensory bridge) for multi-shot coherence
• 🎨 Short Drama Style Quick-Selector: 4-dimension combo system (visual type × render style × color tone × genre)
• 📝 Short drama prompt templates expanded from 1 to 4 variants (CN dialogue / CN voiceover / EN dialogue / EN voiceover)
• 📝 Multi-segment Director's Note template adds transition strategy declaration
• 📝 5 complete short drama examples covering: plot-twist dialogue / voiceover monologue / action conflict / 2D anime / transition strategy
• ✅ 54 tests pass

• 🎬 Smart Multi-Segment Storyboard: Videos >15s auto-split into multiple independent prompts (each ≤15s, min ≥8s)
• 📝 Multi-segment coherence: unified style preamble / three-layer lighting / sound design / transition frames / forbidden items
• 📝 Step 5 adds multi-segment output format template (CN / EN)
• 📝 New 60-second desert Kali/Escrima 4-segment full example
• 🔧 Validation script adds validate_multi_segment() cross-segment consistency check
• ✅ 54 tests pass (incl. 4 new multi-segment validation tests)

• 🌐 Bilingual Prompt Output: Chinese users → Chinese, non-Chinese → English, with auto language detection
• 📝 All structural templates, delivery formats, and multimodal tips now include English versions
• 🛡️ Camera Term Disambiguation (Rule 9): Chinese uses Chinese camera terms, English uses full phrases — avoids Seedance review false positives
• 🔧 Validation adds check_ambiguous_terms() bare-word detection + --lang flag + English word-count length check
• 🔧 New Slow Motion vs Speed Ramp conflict detection
• 🔧 detect_language() expanded with CJK Extension A + full-width punctuation support
• 📚 cinematography.md adds "Seedance Safe Phrasing" column
• ✅ 50 tests pass (incl. bilingual + review safety tests)

• 🎨 Cel-Shaded CG Style: New complete four-axis parameterized entry (distinct from anime's explosive energy — positioned for contemplative narrative)
• 🧱 Anime / NPR Material Library: Anime skin / hair / cartoon metal / cartoon fabric — 4 non-photorealistic materials
• 📦 Anime Game PV Variant: Scenario template adds Cel-Shaded sub-template + ice-attribute character example
• ⚠️ Conflict matrix adds: Cel-Shade vs Realistic PBR material
• 🔧 Validation adds Cel-Shade vs PBR style conflict detection

• 🎬 Camera Upgrade: New focal length narrative psychology, dynamic focus paradigms, physical mount chapter (7 specialty rigs)
• 🎨 Director Styles: New Fincher / Deakins / Kurosawa / Makoto Shinkai + Anime Explosion / Xiaohongshu Aesthetic (incl. de-named safe prompts + forbidden items)
• 💡 Quality Upgrade: Anti-plastic manifesto, film stock library (5 types), material texture library (8 types), lighting combo quick-reference (4 sets), organic imperfection library, quality conflict matrix
• 🎬 Scene Expansion: New Game PV / Horror-Thriller / Travel-City / Pet-Cute / Before-After / Meme-Loop, totaling 16 scenarios + physics damping appendix
• 🎙️ Sound Upgrade: Spatial acoustic modifiers (7 types), material-based onomatopoeia refinement (7 pairs)
• 🔧 Validation Enhancement: Filler word warning → error hard-block, optical physics conflict detection, style conflict matrix, duration-aware time-slicing, 35 tests pass

• 🎉 Initial release
• SKILL.md core workflow
• 6 professional knowledge base files
• Python validation script + test cases
• 20+ director style mappings
• 10 vertical scenario templates

MIT-0 (MIT No Attribution) License