JSON as Art Director: 96 Reusable AI Image Styles

What it does

The AI Visual Prompt Cookbook is a curated library of 96 reusable style.json templates for generating consistent AI imagery. Each template encodes a complete visual system—typography rules, color palettes, composition logic, and negative-space handling—into a structured file you drop into a multimodal LLM. You fill in variables like SUBJECT or MAIN_TEXT, and the model outputs a detailed, style-locked prompt ready for image generation.

The interesting bit

Instead of treating prompts as prose to be rewritten every time, the project treats them like software interfaces: typed variables, conditional layout branches for different aspect ratios, and “style fidelity locks” that prevent the model from drifting. It is essentially prompt engineering reduced to a configuration file.

Key highlights

• 96 distinct visual systems spanning categories like zine collages, type posters, and editorial minimalism, each shipping with two preview images
• Structured variable substitution via environment_variables and examples[*].values keeps the style direction intact across generations
• Optimized for end-to-end multimodal LLMs such as ChatGPT Images 2 and Nano Banana Pro that ingest long JSON and reason before generating
• A separate Copy Prompt Library offers flattened, copy-ready text for quick one-off use
• Online gallery and multilingual README support for browsing outside the repo

Caveats

• Primary support is limited to specific multimodal models; the README notes that other LLMs accepting long JSON “may also work but are not the primary recommendation”
• This is a template collection, not a generation engine—you still need access to capable, likely commercial, image models

Verdict

Worth bookmarking if you generate campaign visuals, thumbnails, or editorial art and want repeatable style systems. Skip it if you are looking for a self-hosted image generator or prefer hand-writing free-form prompts every time.