# GPT-Image2-Skill: a prompt gallery that was turned into two agent skills and a CLI

> A bilingual prompt collection for OpenAI's image models, packaged as installable skills for Claude Code and Codex plus a Python CLI. The interesting part is the packaging, not the pictures.

**wuyoscar/GPT-Image2-Skill** — GPT Image 2/2.5 prompt gallery, image prompt library, agentic skill, and CLI for OpenAI image generation/editing

- Repository: https://github.com/wuyoscar/GPT-Image2-Skill
- Stars: 5,710 · Forks: 472
- Language: Python
- License: MIT
- Published: 2026-10-07 · Updated: 2026-10-07 · Language: en
- Canonical page: https://hysenlabs.com/projects/wuyoscar-gpt-image2-skill

## Three surfaces built from one prompt collection

wuyoscar/GPT-Image2-Skill bills itself as a prompt gallery, an image prompt library, an agentic skill and a CLI, and that list is accurate. It has 5,370 stars and 459 forks, three open issues, and its last push was on 2026-09-09. The language is Python, the license is MIT, and the topics enumerate exactly which use cases the author has in mind: agent-skills, claude-code-skill, codex-skill, image-editing, research-figures, text-to-image and ui-mockups among them.

The surfaces are worth separating because they serve different readers. The README is a browsable gallery for a human deciding what to try. The skills are instructions for an agent runtime. The CLI is for people who want a script.

The project structure makes that split explicit. `skills/` holds the two agent skills with their reference files, `src/` holds the Python package, `docs/` holds assets and guides, `tests/` is present, and `.claude-plugin/` declares plugin metadata for Claude Code. `pyproject.toml`, `CHANGELOG.md`, `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, `SECURITY.md`, `SUPPORT.md` and `TODO.md` sit at the root, which is more governance than most prompt collections bother with.

The README is bilingual, with English and a separate Chinese file linked at the top, and the documentation note in the release history confirms that switching was added deliberately.

## The gallery as numbered, categorised reference material

The README describes the collection as 163 numbered entries across 31 categories. Entries are referenced by number in the surrounding prose, which is what makes them usable as a lookup rather than as wallpaper. Sample A of the showcase is entry No. 113, and sample B builds on the reference from No. 54.

The prompts are long and specific in the way that works for image models. The Meridian 8 example, an exploded-view illustration of a fictional wristwatch, does not say nice or professional; it names the eleven components to separate vertically, the materials to use, the exact in-image strings to render, and the numbered callouts to include.

```text
Create a premium technical exploded-view illustration of a fictional mechanical wristwatch called the Meridian 8, centered on a dark slate background with fine blueprint grid accents.
```

The edit prompt in sample B works the same way, anchoring on a supplied reference image and then specifying what changes: keep the street footprint and the corner cafe recognizable, transform into a three-storey cutaway diorama, open the front-facing walls to reveal the espresso bar and jazz lounge, then list the material and lighting treatment. That structure is a useful thing to copy, because it separates the parts you keep from the parts you change, which is exactly what an edit model needs.

The 2.5 samples share a stated configuration: model `gpt-image-2.5-sunburst`, 2048x2048, quality `high`. Attribution is handled explicitly, with sample B crediting an external community prompt collection and linking the source.

## Two skills and the context-window argument behind them

The v0.2.0 release notes describe the change that matters most for agent use: the gallery was split into category-level `gallery-*.md` files so an agent can load the relevant prompt slice without filling the context window. The earlier arrangement had the whole collection in one place, which works for a human scrolling a page and works badly for a model reading a file.

Two skills ship. One handles image generation and editing, the other image-to-prompt extraction. That second one is the more interesting design idea: given an existing image, extract a prompt that would reproduce it. It turns the collection from a lookup table into something that can be extended with a reference the user already has.

The install story is broader than a single runtime. Topics list claude-code-skill and codex-skill, and the README names Claude Code, Codex, OpenClaw, Hermes Agent and other skill-capable runtimes. The release notes also mention clarifying that `$skill-installer` is a built-in Codex skill-management helper, and adding a manual fallback path for people without it. A repo that ships an agent skill has to solve distribution, and this one tries to work across several agent hosts rather than betting on one.

## The CLI wraps the Image API and nothing more

The Python side is small and honest about its scope. `pyproject.toml` declares the distribution name as `gpt-image-cli`, version 0.2.0, requiring Python 3.11 or newer, with two runtime dependencies: `openai` at 2.32.0 or newer, and `python-dotenv`.

```toml
[project.scripts]
gpt-image = "gpt_image_cli.cli:main"
```

So the CLI is a console entry point over the official SDK, with dotenv support so the API key can live in a local file. The classifiers mark it as beta and target Python 3 only.

The README is specific about what the CLI is for. For editing, you export the relevant PDF pages as PNG, WebP or JPG and attach them with `-i`. It notes that the Image API outputs PNG, JPEG or WebP, which means editable SVG or PowerPoint shapes need a separate authoring step, and that for slide work it is worth aiming at vector-style diagrams, icons and layouts rather than expecting a slide file back.

That caveat is the most useful sentence in the README for anyone planning real work with it. A tool that generated an editable artefact would be a different category of thing, and this one is not claiming to be it.

## Numbers and dates that do not agree across the repository

Three details in this repository contradict each other, and none of them are serious, but each one matters if you cite the project.

First, the count. The README overview table says 163 numbered entries across 31 categories. The v0.2.0 release notes say the README was compacted while keeping the full 162-prompt catalog in the Reference Gallery. The v0.1.0 release, the day before, said 151 curated prompts. So the catalogue grew from 151 to 162 by v0.2.0 and is described as 163 in the README table. If you need an exact number, the README table is the most recent statement, and the release trail explains why the two disagree.

Second, the update date. The README table says last update 2026-09-10. The repository's last push was 2026-09-09. A one-day gap in the wrong direction suggests the README was written in anticipation of a commit, or that the table was hand-edited.

Third, the repository name in URLs. The project is named `GPT-Image2-Skill`, but every link in `pyproject.toml` and in the badge URLs points at `wuyoscar/gpt_image_2_skill`. GitHub redirects those, so they resolve, but the canonical name and the linked name differ in case and separators. The `name` field in the project metadata is a third variant again: `gpt-image-cli`. Nothing is broken; all three simply identify different things, and a reader looking for the code under `src/` should search for the Python package rather than the repository name.

## Conclusion

What makes this repository more than a list of prompts is the decision to split the gallery into per-category reference files so an agent loads only the slice it needs, and to keep a real CLI alongside the skill instructions. The prompts themselves are strong on layout language and weak on the constraints that make image models behave, which is why the author says plainly that showing a reference beats describing at length. If you want the catalogue, browse the README gallery index. If you want it in your agent runtime, install the skill and let it read the category files on demand.

## FAQ

### What is GPT-Image2-Skill and how do I use it?

It is three things from one prompt collection: a browsable gallery of numbered prompts for OpenAI's GPT Image 2 and 2.5 models, two installable agent skills, and a Python CLI installed as the gpt-image console command. The README gives per-category gallery indexes and the skills directory carries the same material for agent runtimes.

### Does this repository contain prompts for GPT-1 or GPT-2 language models?

No. Those are different things entirely. The name refers to OpenAI's GPT Image 2 and 2.5 image generation models, and the collection covers text-to-image generation and reference-image editing rather than any language model.

### Can the CLI edit an image from a PDF?

Yes, indirectly. The README's instruction is to export the relevant PDF pages as PNG, WebP or JPG first and attach them with the -i flag. The Image API outputs PNG, JPEG or WebP only, so anything needing editable vector output needs a separate authoring step.

### How many prompts are in the gallery?

The README overview table states 163 numbered entries across 31 categories. The v0.2.0 release notes refer to a 162-prompt catalog and v0.1.0 to 151, so the count grew across releases and the README figure is the newest statement.

## Sources

- [Issues](https://github.com/wuyoscar/GPT-Image2-Skill/issues)
- [License: MIT](https://github.com/wuyoscar/GPT-Image2-Skill/blob/main/LICENSE)
- [README](https://github.com/wuyoscar/GPT-Image2-Skill/blob/main/README.md)
- [Releases](https://github.com/wuyoscar/GPT-Image2-Skill/releases)
- [wuyoscar/GPT-Image2-Skill on GitHub](https://github.com/wuyoscar/GPT-Image2-Skill)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/wuyoscar-gpt-image2-skill