convez.ai
SkillAutomaçãoIntermediário
irenezhu1999-lab

Cinematic AI Skills

irenezhu1999-lab/cinematic-ai-skills

Engenharia de prompts cinemática contra vídeos genéricos de IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Conjunto 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-skills

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

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.

Claude Code License: MIT Domain Models


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

SkillDescriptionStatus
video-tasteAnti-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-directorProfessional 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

  1. Video Read — declare understanding (theme, platform, tone, shot language, fidelity priority)
  2. Three Dials — set values that drive all decisions (DRAMA / MOTION / FIDELITY)
  3. Beat Sheet — sequential micro-screenplay (Initial → Trigger → Emotion → Reaction → Camera → End → Style)
  4. Six-Layer Structure — shot size → angle → movement → lens → lighting → action
  5. Stanislavski Performance — concrete physical actions, not abstract emotions

Full methodology: video-taste/README.md


Evidence Base

SourceFindingLink
Community benchmarkStructured 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 motionSkywork AI
EvalVerse (academic)45 sub-dimensions, emotional progression is #1 differentiator, expert correlation SRCC +0.74–+0.95arXiv: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


MIT License — Built for directors, DPs, and AI video creators who refuse generic output.

Ver o original no GitHub

Bom para

Relacionados