Auto Video Prompt Reverse Engineer v3.0

Advanced AI video prompt reverse engineering. Input video/screenshot/description → structured prompts + reproduction workflow.

When This Skill Triggers

User provides any of:

• Video link (Bilibili, YouTube, TikTok, Xinpianchang, etc.)
• Video file or screenshot(s)
• Text description of a video's visual style
• Request to analyze, deconstruct, replicate, or reverse-engineer a video

Analysis Rules — Every Shot Must Include

Style Identification Checklist

Must identify ALL that apply:

• Director style: Nolan, Villeneuve, Refn, Deakins, Wes Anderson, Snyder, etc.
• Film genre: Cyberpunk, Atomic Punk, Film Noir, Neo-western, Post-apocalyptic, etc.
• Animation style: Anime, cel-shaded, stop-motion, etc.
• CG style: Photoreal, stylized, low-poly, etc.
• Commercial style: Product hero, lifestyle, fashion, tech reveal
• AI artifacts: Temporal flicker, morphing faces, smooth physics, perfect lighting, etc.
• Color grading: Teal-orange, bleach bypass, film noir, vaporwave, golden hour, kodachrome, etc.
• Lens style: Anamorphic flare, tilt-shift, bokeh characteristics, focal length
• Film stock: Kodak Portra, Fuji Pro, Kodachrome, etc.

Model Estimation

Identify likely AI model(s) used:

• Kling: Smooth motion, Chinese prompt friendly, good physics
• Seedance (小云雀): Audio-visual sync, immersive short film mode, character consistency
• Runway Gen-3: Camera controls (Pan, Zoom, Roll), cinematic quality
• Veo: Advanced cinematographic natural language understanding
• Sora: Narrative prompts, long duration, complex physics
• Pika: Short clips, motion parameter control
• SVD: Image-to-video, motion_bucket_id control
• Wan/CogVideoX: Chinese-optimized, shorter clips
• Midjourney/Flux: Keyframe generation (not video)

Prompt Reverse Engineering Template

For EACH shot, output:

Shot XX

Content: [what's happening]
Camera Language: [shot type + composition]
Motion: [camera movement type]
Lighting: [lighting setup]
Color: [palette + temperature]
Material/Texture: [film grain / CGI / realistic]
Subject Action: [what the subject does]

Positive Prompt: [structured prompt: Subject + Action + Environment + Camera + Lighting + Color Grade + Style + Technical] Negative Prompt: [what to exclude] Camera Prompt: [lens focal length, movement, angle] Style Prompt: [director reference, film stock, genre] Lighting Prompt: [specific lighting setup] Parameters: [aspect ratio, FPS, motion scale, CFG, model]

Output Format

Use this structure for every analysis:

`

Video Overall Style Analysis

Video Type

• [Short film / Commercial / MV / Documentary / etc.]

Overall Style

• [Atomic Punk + Post-apocalyptic Western / Cyberpunk / etc.]

Director Reference

• [Most similar director(s) and why]

Editing Rhythm

• [Slow build / Fast cuts / Montage / etc.]

Estimated AI Model(s)

• [Primary model + supporting tools]

AI Generation Artifacts Detected

• [List specific tells]

Shot Breakdown

Shot 01

[Full analysis per template above]

Shot 02

[Continue for ALL key shots]

Global Reverse-Engineered Prompts

Global Style Prompt (applies to all shots)

[Master style anchor prompt]

Global Negative Prompt

[What to always exclude]

Global Camera Prompt

[Default lens, movement vocabulary, aspect ratio]

Parameter Estimation

Reproduction Workflow

1. Keyframe Generation → Midjourney / Flux → Generate character concept art + scene reference images
2. Video Generation → Choose based on style:
   • Ads / Product / Tech: HappyHorse (快马) — Best for advertising & futuristic tech style
   • Narrative / Short Film: Kling / Seedance / Runway → Text + reference image to video
   • Creative / Art: Veo / Sora → Complex cinematic scenes
   • Quick iteration: Pika / SVD → Short clips, rapid testing
3. Director Method → Write director-style prompts (WHY characters do things, not just WHAT)
4. Audio Sync → Seedance immersive mode for audio-visual sync / manual foley / HappyHorse dialogue support
5. Color Grade → DaVinci Resolve → Match LUT, cascade correction
6. Enhance → Topaz Video AI → Upscale + denoise + stabilize
7. Compose → Premiere / Final Cut → Edit, rhythm cuts, music sync
8. Final Pass → Film grain overlay, letterboxing, sound mix

HappyHorse Prompt Optimization Rules

When generating HappyHorse prompts, apply these transformations:

• Remove negatives: Replace "no helmet" → "helmet removed, resting on metal stand beside"
• Visual substitution: Replace "back to camera" → describe what the back LOOKS LIKE (helmet top reflecting light, shoulder armor V-shape)
• Three-level shot control: [Shot type] + [Angle anchor] + [1-2 micro details]
• Cinematic keywords: Always append quality tags (cinematic quality, film grain, etc.)
• Chinese content: Write prompts in Chinese for best results
• Prompt length: 50-150 characters optimal; too long causes semantic drift
• No dialogue by default: Only add dialogue when user explicitly requests it
• Dialogue duration estimation: Slow speech ≈ 3-4 chars/sec, normal ≈ 5-6 chars/sec; always leave 1-2s buffer for transitions `

Pro Tips (from real creators)

• Tell AI WHY, not just WHAT: "Character presses hat down because wind might blow it off" > "Character presses hat"
• Use non-human characters to bypass uncanny valley (robots, pixels, masks)
• Don't draw storyboards: Use scene reference image + character image + text direction
• Keep AI surprises: When generation gives unexpected good results, fold them into narrative
• Credit your models: List all AI tools used, treat them as "cast and crew"
• Multiple takes: Generate 10-20 variations per shot, select the best action/rhythm

References

See references/model_params.md for model parameters, lens focal lengths, and color grading keyword reference.