What it does

This repo collects 470+ reverse-engineered GPT-Image2 examples and distills them into 20+ industrial prompt templates. The pitch is “Prompt as Code” — compressing prose-style descriptions into atomic, composable schemas (subject, lighting, materials, layout) that agents and scripts can reuse for batch generation or production workflows.

The interesting bit

The project treats prompt engineering as interface design rather than creative writing. It splits prompts into modular parts — components, hierarchy, texture — and ships an Agent Skill file so LLM-based tools can programmatically assemble image requests. There’s also a browsable web gallery where you can preview outputs, copy full prompts, and filter by scenario.

Key highlights

• 470+ cases across 12 categories: UI, infographics, posters, product shots, photography, illustration, characters, architecture, and more
• Atomic schema approach: subjects, lighting, materials, layout as separable, reusable blocks
• Agent-ready: includes a SKILL.md file for LLM agents to consume the style library programmatically
• Live gallery site at gpt-image2.canghe.ai with preview, copy-paste, and generation testing
• MIT licensed; actively maintained with irregular batch updates

Caveats

• The English README is a navigation layer; actual template bodies live in separate docs, so browsing requires some clicking around
• Sponsorship messaging and WeChat community integration are prominent throughout
• “Updated irregularly” — no fixed release cadence

Verdict

Worth bookmarking if you’re building automation around GPT-Image2 or need production-grade prompt structure. Skip it if you’re looking for a casual prompt cookbook or a fully programmatic API wrapper.