What it does

This repo is three things stacked together: a curated prompt gallery for OpenAI’s GPT Image 2, an installable agent skill for Claude Code / Codex / OpenClaw, and a Python CLI (gpt-image) that wraps the image generation and editing endpoints. You can browse prompts for research figures, game HUDs, anime panels, or brand identity kits, then run them immediately without hand-rolling API calls.

The interesting bit

Most prompt galleries are static markdown. This one treats prompts as executable artifacts — each gallery entry ships with just the prompt plus a metadata line for size and quality, and the CLI assembles the actual API call. The skill packaging is unusually thorough: it supports plugin install for Claude Code, $skill-installer for Codex, manual symlinking for Hermes Agent, and uvx/uv tool for standalone CLI use. The author also explicitly warns against dropping generated figures directly into academic papers, which is a refreshing dose of caution in a hype-heavy space.

Key highlights

• Curated, not scraped: the gallery is intentionally small, covering niches like scientific diagrams, tattoo design, typography posters, and UI mockups
• CLI supports text-to-image, multi-reference edits, and mask-based inpainting with sensible shortcuts (--size portrait, --quality low/medium/high as a budget dial)
• Agent skill installs into multiple runtimes without forcing global/shared installs by default
• Reads OPENAI_API_KEY from env, .env, or ~/.env without clobbering already-set variables
• Includes a “budget / quality guide” that maps low to cheap drafts and high to text-heavy final assets

Caveats

• The README notes that gpt-image-2 rejects the --input-fidelity parameter, so the CLI silently drops it; this is a minor rough edge from model version drift
• The gallery is described as “small but mighty” — if you need exhaustive coverage, this is not it
• The author mentions an “automated patch/update workflow behind it” that may be shared later, so some operational details remain opaque

Verdict

Worth a look if you generate images regularly from an agent runtime or terminal and want vetted prompts without the Pinterest scroll. Skip it if you already have a polished internal pipeline or need a massive prompt database.