GPT Image Skill

Generate and edit GPT images from Codex, Claude Code, or another compatible local agent through the user's ChatGPT subscription. Direct prompts stay unchanged; when the user delegates several different designs, the agent develops a distinct image-ready prompt for each concept. Real reference files pass into generation, results stay in the active project, and ready outputs run with bounded parallelism.

install skill → Sign in with ChatGPT → direct prompt or delegated concept prompts + local image inputs
              → single generate or bounded parallel batch
              → built-in $imagegen → <project>/generated-images/*.png

This repository does not call the OpenAI Images API and does not create a separately billed Images API request. Built-in image generation still consumes included ChatGPT/Codex usage and remains subject to plan and workspace limits. OpenAI notes that image generations use included limits 3–5× faster on average than similar non-image turns, depending on quality and size, so parallel batches should stay intentional and small.

Install by pasting one prompt into an agent

Paste this into Codex, Claude Code, or another local coding agent:

Install and verify GPT Image Skill for the current user from:

https://github.com/GENEXIS-AI/gpt-image-skill

For this task, I authorize read-only environment checks; a persistent clone or safe fast-forward update;
user-level installation of missing Git, a supported Node.js 22+ LTS, and Codex CLI;
creation of the gpt-image links for Codex and Claude Code; and starting Sign in with ChatGPT device authorization.

Read AGENT_INSTALL.md at the repository root and follow it as the one-time installation contract.
Do not use the Images API, OPENAI_API_KEY, or API-key login. Do not generate a live image yet.
Pause only if administrator privileges are required, an unrelated existing path would be changed,
local changes would be discarded, or existing Codex authentication would need to be replaced.
Otherwise, install the required components, run bootstrap --target all --yes --json,
and continue until bootstrap's consolidated readiness report shows best_practice_pass=true.
Do not add a separate doctor, plan, inspect, or no-image generation check when bootstrap passes.
Finally, report the persistent clone path, both installed skill paths, and ChatGPT-auth evidence.
Then give me the brief getting_started guide in my language: common aspect ratios,
quality phrases, one creation example, and one reference or revision example.
Do not use unexplained jargon such as "dry-run"; call it a setup check that does not create an image.

This prompt authorizes ordinary user-level setup without authorizing administrator elevation, destructive changes, replacement of existing authentication, a live generation, or a GitHub Star. The full boundary is in AGENT_INSTALL.md.

After setup, invoke $gpt-image in Codex or /gpt-image in Claude Code. The host loads the concise skill only for image tasks; it does not need to reread this README on every request.

What the agent shows after installation

The installing agent should end with a small guide like this, translated into the user's language:

GPT Image Skill is ready. No image was generated during setup.

Common aspect-ratio requests: 1:1, 16:9, 9:16, 4:3, 3:4
Quality phrases: draft, high quality, high detail / final quality

Try:
$gpt-image Create a cozy reading room at sunset, 16:9, high quality.
/gpt-image Use @references/character.png as the character reference and place it in a rainy city, 9:16, high quality.

These are common natural-language requests, not a fixed API size list. Other framing or dimension requests can be written normally, and exact pixel dimensions may vary with built-in image generation. The guide appears once after installation rather than after every image.

Design principles

1. The prompt is authoritative—and so is delegated creative intent

For one direct image or edit, the skill forwards the user's image request unchanged. It does not “improve” a short prompt with unsolicited details. A request such as “make five different poster designs” is different: it explicitly asks the agent to develop five creative concepts. The agent preserves the shared subject, references, brand, text, ratio, and other constraints, then sends one complete, meaningfully different image prompt per output. Job numbers remain in IDs and filenames, never in image prompts.

2. References are files, not descriptions

The bridge requires a readable local PNG, JPEG, or WebP path for each reference. It passes those files into the actual $imagegen call. It never replaces an unresolved reference with a text description and continues anyway.

In Claude Code:

• Prefer @path/to/image.png or an explicit filesystem path.
• A pasted or dragged image visible in Claude is not automatically inherited by a nested codex exec process.
• If Claude exposes an exact readable temporary attachment path, the agent copies that exact file into <project>/generated-images/inputs/ and uses the copy.
• If Claude exposes no path, save the image inside the project and provide that path before generation.
• The skill does not guess from ~/.claude/image-cache; choosing “the newest image” could select the wrong or private file.

3. Revisions always edit the latest result

Each bridge invocation is ephemeral. For “change the result you just made,” the agent must use the previous generated output as the next --edit-target and reattach every still-needed reference. Reusing the original source is a different operation and loses the prior edit.

4. Normal generation stays light

The default path is:

quick ChatGPT-auth check → one generation → minimal PNG sanity check → PATH + inline Markdown

Planning, the setup check that does not create an image (--dry-run), capabilities --json, inspect --input, and detailed JSON remain available for troubleshooting. They are not required before a normal image request.

5. Parallelism is explicit and bounded

generate remains the one-image happy path. For two or more outputs, the skill automatically batches every job whose inputs already exist. Different design concepts run independently; same-design variants may run together while reading one shared edit target or design reference. Only an output-to-input dependency creates another stage. The batch checks ChatGPT auth once and uses default concurrency 2, maximum 4, with no Doctor, planning, inspection, or automatic retries per job.

Features

• Uses Codex's built-in $imagegen under Sign in with ChatGPT
• Uses a host-native image_gen tool directly when the calling host already exposes one
• Blocks OPENAI_API_KEY, API-key Codex login, and Images API fallback
• Installs the same gpt-image skill for Codex and Claude Code
• Preserves direct prompts verbatim and honors delegated multi-concept design work
• Supports one or multiple references with deterministic attachment order
• Supports existing-image edits, follow-up revisions, variations, compositing, transparency, exact text, and dense-layout drafts
• Supports bounded parallel generation for independent concepts and shared-anchor variations
• Saves only inside the active workspace and avoids overwrite by default
• Returns PATH=... and absolute MARKDOWN=... after normal generation
• Supports macOS, Linux, native Windows, and WSL2

Workflow map

The edit target is Image 1. Supporting references follow in command-line order.

Requirements and environments

• Node.js 22 or newer; the current supported LTS is recommended.
• Git when installing from GitHub.
• Codex CLI signed in with ChatGPT, unless the calling host provides native image_gen.
• A ChatGPT/Codex plan and workspace that permit image generation.

Manual installation

Use a persistent clone because the installed skill links point to it.

macOS, Linux, and WSL2

REPOSITORY_URL="https://github.com/GENEXIS-AI/gpt-image-skill"
INSTALL_DIR="${XDG_DATA_HOME:-$HOME/.local/share}/gpt-image-skill"

git clone "$REPOSITORY_URL" "$INSTALL_DIR"
cd "$INSTALL_DIR"
node ./gpt-image/scripts/validate_skill.mjs
node ./gpt-image/scripts/gpt_image.mjs bootstrap --target all --yes --json

Keep WSL2 clones under the Linux h