smixs/visual-skills

What it solves

It transforms AI agents into a professional film crew by moving beyond simple adjective-based prompting. Instead of generic descriptions, it provides a framework for "dramaturgy"—teaching agents how to handle staging, camera movement, lighting, and editing to create cinematic AI video and images that follow real filmmaking principles.

How it works

The project consists of two "Claude Skills" (video and image) that act as a router for a series of reference files the agent must load in a specific order:

1. Dramaturgy: Establishes the scene formula (desire, obstacle, geometry, gaze, rhythm) and editing rules.
2. Universal Rules: Applies general prompting discipline (e.g., show-don't-tell, character anchors).
3. Model-Specific Files: Applies the exact syntax for specific models like Seedance, Kling, Veo, Nano Banana, or GPT Image.
4. Task Modules: Provides specialized vocabulary for camera, lighting, and genre patterns.

Every output is gated by a six-point dramaturgy check and a three-detail audit to ensure the prompt contains physical facts rather than vague adjectives.

Who it’s for

AI filmmakers, content creators, and prompt engineers who want to generate high-quality, consistent AI video and images using professional cinematic direction rather than trial-and-error prompting.

Highlights

• Cinema-First Approach: Prioritizes dramaturgy (emotion, story, rhythm) over model syntax.
• Multi-Model Support: Dedicated syntax files for Seedance, Kling, Veo, Nano Banana, and GPT Image, with universal rules for Runway, Luma, Pika, and Sora.
• Integrated Pipeline: Chains with a creative-director skill for a full workflow from idea/script to keyframes and final motion.
• Strict Quality Gating: Mandatory audits that prevent the agent from returning prompts that rely on banned words like "cinematic" or "epic."