UltraReal Prompt Architect

An installable Agent Skill that turns a short image idea into a production-ready photorealistic image-generation prompt.

It writes the way a photographer or cinematographer briefs a shot: subject, place, action, motivated light, honest materials, and a camera choice that matches the job. It does not dump 8K, masterpiece, ultra detailed, cinematic lighting onto every request.

Core principle: Don't describe "AI realism." Describe reality.

What it does

• Expands one-line ideas (a man riding a Royal Enfield in Ladakh) into a specific, physically believable photograph brief
• Preserves the user's subject, clothing, location, culture, action, and objects
• Infers only supporting photographic facts (light, weather, lens, wear)
• Adapts dialect for Midjourney, Flux, Stable Diffusion / SDXL, GPT Image, Gemini / Nano Banana, Ideogram, Firefly
• Supports styles and mixes: documentary, cinematic, fashion, street, travel, product, architecture, wildlife, historical, religious/devotional, editorial
• Refines existing prompts (more realistic, make it night, 9:16, less AI, video-ready) without rewriting the story
• Emits a short negative only when the named model benefits from one

Install

This package follows the Agent Skills standard: a folder named ultrareal-prompt-architect containing SKILL.md.

Claude Code / compatible agents

# user-level
cp -R ultrareal-prompt-architect ~/.claude/skills/ultrareal-prompt-architect

# or project-level
cp -R ultrareal-prompt-architect .claude/skills/ultrareal-prompt-architect

Claude.ai — zip the ultrareal-prompt-architect folder and upload it under Settings → Capabilities → Skills.

Other agents (Cursor, Codex, OpenCode, etc.) — copy the folder into that product's skills directory. The only required file is SKILL.md; references/ load on demand.

The folder name must stay ultrareal-prompt-architect to match the name field in the frontmatter.

How an agent should use it

1. The description in SKILL.md is the trigger (image prompts, photorealism, named models, "less AI", etc.).
2. On trigger, load SKILL.md and follow the workflow.
3. Open files in references/ only when that engine is needed (faces, a named model, food, etc.).
4. Return the output format in SKILL.md: Main Prompt, optional Negative, Suggested Settings.

You can also invoke it explicitly:

/ultrareal-prompt-architect a ceramic mug on a teak table, morning window light

Package layout

ultrareal-prompt-architect/
├── SKILL.md                         # workflow, engines, output contract
├── README.md
├── LICENSE
└── references/
    ├── principles.md                # research distilled into rules
    ├── cameras-lenses.md
    ├── lighting.md
    ├── composition.md
    ├── human-realism.md
    ├── materials.md
    ├── subject-playbooks.md
    ├── model-adaptation.md
    ├── style-presets.md
    ├── anti-ai-look.md
    └── examples.md

Example

Input: A man riding a Royal Enfield in Ladakh

Main Prompt (abridged): a travel photograph of a man on a dust-coated Royal Enfield in a Ladakh curve; late-morning high-altitude sun from camera right; 35mm at f/5.6; broken tarmac, pale dust, barren ridges; unretouched documentary colour.

Full worked cases — including Shiv Ji in the Himalayas, product, food, architecture, and model-specific ports — are in references/examples.md.

What it will not do

• Generate the image unless you also ask the host agent to generate
• Invent undocumented model parameters
• Overwrite identity, sacred iconography, or specified wardrobe
• Optimize for illustration, anime, or 3D unless you change the brief

License

MIT. See LICENSE.