🎨 Visual Skills for Claude — Image & Video Prompting

🇷🇺 Читать на русском

Two professional Claude Skills for AI visual content production. They write production-grade prompts for the leading image and video models — picking the right model for the task, applying its specific syntax, and returning a copy-paste-ready prompt.

This is what a creative director, copywriter, or AI-content team uses instead of "be cinematic, 4k, masterpiece" filler.

✨ Supported Models

🖼️ Image generation models

🎬 Video generation models

🤝 Compatible With

These are plain Claude Skills — markdown files plus a packaged .skill archive. They work in any agent or IDE that supports the Claude Skill format:

The skills work with Claude Opus, Sonnet, Haiku, and degrade gracefully on GPT / Gemini / open-weights agents (the markdown is model-agnostic).

📦 What's in the Repo

visual-skills/
├── image/              # Source folder for the image-prompting skill
├── image.skill         # Packaged skill — drop-in installer
├── video/              # Source folder for the video-prompting skill
├── video.skill         # Packaged skill — drop-in installer
├── README.md / README.ru.md
└── LICENSE             # MIT

🖼️ image — What It Does

Writes prompts for AI image generation. Picks Nano Banana or GPT Image 2 based on the task, applies the model's specific syntax, returns a copy-paste-ready prompt with a header (model, quality, size).

Tasks covered:

• 📰 Editorial photography, posters, ad creatives
• 🛍️ Product shots, packaging, mockups
• 🖥️ UI mockups and product screenshots
• 📊 Infographics, diagrams, slides
• ✏️ Edits — try-on, lighting/weather swap, object removal, restoration, localization
• 👤 Character continuity across multiple images
• 🎞️ Storyboards, comics, sequential narrative
• 📐 Sketch-to-photo, wireframes, 2D-to-3D, floor plans

Model split:

Reference files inside image/: models.md, nano-banana.md, gpt-image.md, golden-rules.md, prompt-framework.md, creative-direction.md, text-rendering.md, editing.md, characters.md, slides.md, storyboards.md, structural.md, dimensional.md.

🎬 video — What It Does

Writes prompts for AI video generation. Operates as a hybrid Director / Screenwriter / Editor — applies cinematic dramaturgy (scene formula, Murch Rule of Six, blocking, staging) and the model-specific syntax (Seedance multi-shot, Kling Element Binding, Veo JSON / dialogue).

Tasks covered:

• 🎯 Single 5-second clips and stitched multi-clip stories (15s / 30s / 60s+)
• 🎞️ Director treatments and shot lists (14-field shot card)
• 📋 Storyboards from script
• 🔧 Prompt audits ("here's my prompt, fix it")
• 📝 Translating scripts and storylines into shot-by-shot prompts
• 🔗 Continuity across clips (character lock, wardrobe, lighting logic)
• 🎭 Genre patterns: commercial, music video, drama, action, fashion, UGC, product film

Model split:

Reference files inside video/: dramaturgy.md, universal-rules.md, seedance.md, kling.md, veo.md, role-modes.md, patterns-and-genres.md, camera-lighting-vocabulary.md, fixes-and-skeletons.md.

🚀 Installation

Option A — Install the packaged .skill

Download image.skill and/or video.skill from this repo and load through your Claude client:

# Claude Code
claude install image.skill
claude install video.skill

Option B — Clone the source

git clone https://github.com/smixs/visual-skills.git

Then copy the image/ and/or video/ folders into your skills directory:

# Claude Code
cp -r visual-skills/image  ~/.claude/skills/
cp -r visual-skills/video  ~/.claude/skills/

# Cursor / Windsurf — copy into your project's rules folder
cp -r visual-skills/image  .cursor/rules/

💡 Usage Examples

Image — quick prompts:

"Сделай промпт для постера офисной кружки с надписью BEST DAY EVER, фон #f5f5dc, 16:9"

"Edit this product shot — change the background to plain white, keep the bottle exactly as is"

Image — model-aware:

"Use GPT Image 2 to mock up a Spotify-like UI for a meditation app, quality high"

"Use Nano Banana Pro — cinematic photograph of the Charles Bridge in Prague at golden hour, must be architecturally accurate"

Video — single prompt:

"Напиши промпт для Seedance — голодный мужик ночью находит последнюю сосиску в холодильнике, 5 секунд, мульти-шот"

Video — full breakdown:

"Раскадруй 30-секундный ролик про чувство вины. Главная эмоция — guilt. Опорный объект — телефон с непрочитанным сообщением."

"Audit this prompt: [...]. What's broken, how to fix?"

"Translate this script into 6 × 5-second Seedance prompts."

How It Works (short)

Each SKILL.md is a thin router. The body says "before producing any prompt, load these reference files in this exact order". The actual rules — model-specific syntax, dramaturgy, the Details Law, banned phrases that hurt the model — live only in references/. This forces the agent into the references and prevents lazy generic output.

For video specifically, every shot must own three concrete details: environmental pressure (cold refrigerator light, wet asphalt, flickering fluorescent), physical micro-action (jaw locks, knuckles whiten), and a sound or visual motif. Words like "cinematic", "epic", "stunning", "masterpiece" are banned — they don't render.

Credits & Sources

• Nano Banana — Google Gemini 3 Pro Image / Flash Image, prompting via fal.ai and Google AI Studio guides.
• GPT Image 2 — OpenAI, via OpenAI's developers cookbook and fal.ai's GPT Image 2 prompting guide.
• Seedance — ByteDance Seed, official Seedance 2.0 docs.
• Kling — Kuaishou, official Kling docs.
• Veo — Google DeepMind, official Veo docs.
• Video dramaturgy — Walter Murch (In the Blink of an Eye, Rule of Six), Akira Kurosawa (environment as character), David Fincher (motivated camera), Steven Spielberg (spatial clarity), Jonathan Glazer (one-sentence music video), Bong Joon Ho (storyboarding after locations).

License

MIT — fork it, adapt it, ship better visual content.

Tags: claude · claude-skills · claude-code · claude-agent-sdk · prompt-engineering · ai-image-generation · ai-video-generation · nano-banana · gpt-image · gpt-image-2 · seedance · kling · veo · creative-director · cursor · windsurf · cline · opencode · hermes-agent