Cinematic AI Skills
irenezhu1999-lab/cinematic-ai-skills
Engenharia de prompts cinemática contra vídeos genéricos de IA
Ver no GitHubConjunto de skills de produção para geração de vídeos AI com qualidade cinemática. Utiliza estruturas de prompt baseadas em direção de fotografia, composição cinematográfica e pesquisa acadêmica para eliminar artefatos comuns de vídeos gerados por IA (expressões congeladas, iluminação plana). Reduz taxa de retry em 75% através de prompts estruturados.
Instalação
npx convez-ai irenezhu1999-lab/cinematic-ai-skills1 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
- Geração de vídeos comerciais e criativos com qualidade profissional
- Produção de conteúdo audiovisual para redes sociais (TikTok, Douyin) com direcionais específicos
- Briefing estruturado para modelos de vídeo (Sora, Kling, Wan2.1) com especificações cinematográficas
Veredito

Skill bem fundamentada que trata geração de vídeo como direção profissional, com validação em benchmarks comunitários e acadêmicos (EvalVerse). Particularmente útil para quem quer escapar do padrão visual genérico de modelos de IA; requer engajamento ativo com conceitos cinematográficos.
README
A tradução para o português deste README ainda está na fila.
cinematic-ai-skills
Production-ready Claude Code skills for AI video generation.
Stop generating generic AI-looking videos — apply directorial prompt structures backed by community benchmarks, platform guidelines, and academic research.
Why This Exists
AI video models default to the same traps:
- ❌ Frontal stare + frozen smile
- ❌ Flat even lighting, gradient backgrounds
- ❌ Vague adjectives the model can't interpret ("beautiful", "epic")
- ❌ Simultaneous actions causing distortions
These skills fix that by treating video generation like briefing a DP and actor on set — you specify shot type, camera movement, lighting direction, action beats, and emotional arc.
Evidence: Community benchmark shows structured prompts reduce retry rate from 6.4 to 1.8 (75% reduction). Sora 2 official guidelines and EvalVerse academic benchmark (34 filmmakers, SRCC +0.74–+0.95) validate the same principles.
Skills
| Skill | Description | Status |
|---|---|---|
| video-taste | Anti-slop cinematic prompt engineering• Video Read intent inference• Three dials: DRAMA / MOTION / FIDELITY• Beat Sheet (7-step micro-screenplay)• Six-layer directorial structure• Emotion-visual mapping | ✅ v1.1.0 |
| cinematography-director | Professional cinematography knowledge base• Shot language vocabulary• Performance direction (Stanislavski)• Composition law library• Lighting setup templates | 🚧 Coming soon |
Quick Start
Install
/plugin marketplace add irenezhu1999-lab/cinematic-ai-skills
/plugin install video-taste@cinematic-ai-skills
Example
Before (generic prompt):
A magical video of a girl receiving a Hogwarts letter. Make it emotional and epic.
→ Model outputs: frontal stare, flat lighting, vague "surprised" expression
After (video-taste structured prompt):
Video Read: Harry Potter letter arrival, Douyin audience,
emotional tone: surprise + sacred, shot language: slow push-in + volumetric particles,
facial fidelity: high.
Dials: DRAMA 8 / MOTION 5 / FIDELITY 9
Close-up, eye-level, slow push-in from medium to close-up.
85mm portrait lens, shallow DOF. Warm amber rim backlight intensifying.
Subject holds envelope; notices Hogwarts crest and pauses;
eyes widen and mouth opens; breaks into surprised smile
as golden particles drift left to right.
→ Model outputs: directional lighting, sequential action beats, natural expression arc
Core Methodology
The insight: AI video models are time-series generators. They need what happens at t1, t2, t3 — not what the vibe should be.
video-taste approach
- Video Read — declare understanding (theme, platform, tone, shot language, fidelity priority)
- Three Dials — set values that drive all decisions (DRAMA / MOTION / FIDELITY)
- Beat Sheet — sequential micro-screenplay (Initial → Trigger → Emotion → Reaction → Camera → End → Style)
- Six-Layer Structure — shot size → angle → movement → lens → lighting → action
- Stanislavski Performance — concrete physical actions, not abstract emotions
Full methodology: video-taste/README.md
Evidence Base
| Source | Finding | Link |
|---|---|---|
| Community benchmark | Structured prompts: 1.8 retries avg vs 6.4 for vague prompts (75% reduction) | Cyber Corsairs |
| Sora 2 guidelines | "One camera idea per shot", physical details reduce uncanny motion | Skywork AI |
| EvalVerse (academic) | 45 sub-dimensions, emotional progression is #1 differentiator, expert correlation SRCC +0.74–+0.95 | arXiv:2605.23271 |
Roadmap
- video-taste v1.1.0 — Beat Sheet + six-layer directorial structure
- cinematography-director v1.0.0 — shot language dictionary, composition laws, lighting templates
- Integration examples — Runway, Pika, Claude API workflows
- Multi-shot sequencing extension (when T2V models support)
- Audio sync layer (when audio-capable models mature)
Contributing
Open to:
- Video Read examples for the example library
- Emotion-visual mapping table expansions
- Bug reports and prompt structure improvements
Citation
If you use these skills in research or production:
@software{video_taste_2026,
title = {video-taste: Anti-slop AI video generation skill},
author = {Irene Zhu and Claude Sonnet 4.6},
year = {2026},
url = {https://github.com/irenezhu1999-lab/cinematic-ai-skills}
}
Credits
- video-taste adapted from Leonxlnx/taste-skill
- Beat Sheet methodology from Cyber Corsairs
- Evaluation framework aligned with EvalVerse
MIT License — Built for directors, DPs, and AI video creators who refuse generic output.