gpt-image-bridge — скилл

---
name: gpt-image-bridge
description: Use when the user asks for an image, mockup, logo, avatar, hero image, illustration, diagram, visual reference, or any generated picture, or when a design skill needs an image produced. Bridges to OpenAI's GPT Image 2.5 through the codex CLI using a ChatGPT subscription — no API key.
---

# gpt-image-bridge

This skill adds image generation to Claude Code by shelling out to the `codex` CLI and its `image_generation` tool. OpenAI's [Images 2.5 announcement](https://openai.com/index/introducing-chatgpt-images-2-5/) includes availability in Codex. Codex is authenticated via the user's ChatGPT subscription, so no API key is required.

The bridge uses the image model provided by Codex; it does not select or verify a particular backend model. The executable name `gpt-image-2` is retained for compatibility and does not pin the model to GPT Image 2. Do not pass an image-model ID to `codex --model`: that option selects the coding model, not the image backend.

## When to use

Invoke the wrapper any time the user wants an image produced. Examples:

- "Generate a hero image for..."
- "Make me a mockup of..."
- "I need an avatar / logo / illustration / diagram of..."
- "Show me what X would look like"
- A design-taste skill (e.g. `image-taste-frontend` from [Leonxlnx/taste-skill](https://github.com/Leonxlnx/taste-skill)) is driving an image-first workflow and needs an image generated

Do **not** invoke it for:

- Small programmatic PNGs (1px dots, solid-color placeholders, charts) — write those directly with a few lines of Python.
- Editing an existing image pixel-by-pixel.

## How to call it

```bash
~/.claude/skills/gpt-image-bridge/bin/gpt-image-2 "<prompt>" <absolute-output-path.png> [--size WxH]
```

- **Prompt** should be dense and art-directed: composition, lighting, camera/lens, mood, style reference. Terse prompts produce generic output.
- **Output path** must be absolute. `/tmp/` works for throwaways; a project-local `design/` directory for kept assets.
- **Size is optional** — if omitted, the model chooses its own dimensions. Only pass `--size` when the user specifies one or the layout requires a particular aspect ratio.
- Calls routinely take 4–6 minutes (codex reasons before calling the image tool; exact latency depends on the user's codex `reasoning_effort` config). Set the `Bash` tool timeout to the **maximum, 600000 ms**, or run it in the background and poll.
- The wrapper prints the absolute output path to stdout on success, or a tail of the codex log to stderr on failure.

After the wrapper returns, `Read` the PNG back into context so you can analyze it before coding or responding.

## Under the hood

The wrapper runs:

```