gongnyang/gongnyang-prompt-kit

What it solves

This project provides a structured system for transforming vague, single-sentence image generation requests into high-quality, production-ready prompts for the gpt-image-2 model. It prevents common pitfalls like "naive prompts" (which lead to generic results) and the use of outdated Stable Diffusion (SD) terminology that creates noise in modern models.

How it works

The kit operates as a "compiler" for prompts. It uses a routing table to map a user's rough request to a specific professional pattern (e.g., cinematic key art, typo-posters, or editorial fashion) and then pulls detailed reference files to build a precise prompt.

Key mechanisms include:

• Routing Table: Maps requests to specific categories (C1-C12), typo-poster patterns (TP1-TP17), or promotional materials (P1-P12).
• Validator Gate: A Node.js script (check_prompt.mjs) that checks prompts against a whitelist of allowed terms, ensuring no forbidden negative prompts or incorrect aspect ratios are used before generation.
• Strict Rules: Enforces the use of positive descriptions instead of negative prompts, specific HEX color palettes, and result-based descriptions (e.g., "shallow depth of field" instead of naming a specific camera model).

Who it’s for

Users of Claude Code and gpt-image-2 who want professional-grade visual assets (posters, key art, editorial photos) without manually engineering complex prompts.

Highlights

• Comprehensive Library: Includes a vast array of presets for cinematic key art, typography layouts, and promotional design patterns.
• ** uma Validator**: A built-in tool to catch "SD-style" noise and size-lock violations before they reach the model.
• Professional Workflow: Integrates directly into Claude Code as a skill, allowing for a seamless transition from rough idea to validated prompt.
• Result-Oriented Prompting: Focuses on concrete visual descriptions and numerical values (HEX, Kelvin) rather than vague adjectives.