ChatGPT Image (gpt-image-2)

Image generation skill built on OpenAI's gpt-image-2. Optimized for three high-leverage use cases:
infographics, meta-ad creative, and viral LinkedIn images. Multi-image composition is the
distinguishing capability — pass 2 or more reference images to produce one combined scene.

When to use this skill

• Editorial / magazine / poster layouts — typography precision is gpt-image-2's strength
• Multi-reference compositing — combine product + lifestyle + brand assets in one image
• Text-heavy infographics — stats, headlines, sections rendered legibly inside the image
• Meta ad creative — hook overlay + product shot + lifestyle context in a single composition
• Viral LinkedIn scroll-stops — bold typographic posters with high contrast

When NOT to use this skill

• Multi-slide carousel decks → gpt-image-2 generates one image at a time and cannot hold layout continuity across slides
• Realtime / streaming generation → not supported by gpt-image-2
• Transparent PNGs → not supported by gpt-image-2

Setup

Requires OPENAI_API_KEY environment variable and OpenAI organization verification.
See README.md for full setup. One-time install:

cd ~/.claude/skills/chatgpt-image
npm install

Single-shot generation

node ~/.claude/skills/chatgpt-image/tools/generate.js \
  --prompt "Editorial magazine poster: 'Greater Precision and Control'. Bold serif typography, geometric shapes in black, red, cream. Bauhaus-inspired." \
  --size 1024x1536 \
  --quality high \
  --output ./poster.png

Multi-image composition (the killer feature)

Pass --reference-image multiple times to combine references into one scene:

node ~/.claude/skills/chatgpt-image/tools/generate.js \
  --prompt "A premium gift basket on a white studio background with the items from the reference images arranged inside. Add a ribbon and 'Relax & Unwind' label in handwritten script." \
  --reference-image ./body-lotion.png \
  --reference-image ./candle.png \
  --reference-image ./soap.png \
  --output ./giftbasket.png

When references are provided, the tool calls the Edits endpoint. Without references, it
calls the Generations endpoint. Same flag surface, different backend.

CLI options

--prompt "<text>"               Required. Image description.
--reference-image <path>        Repeatable. Adds a reference image. Triggers Edits endpoint.
--output <path>                 Output path. Default: ./output.png
--size <size>                   Any WxH where max edge ≤ 3840px, both edges multiples of 16,
                                long:short ratio ≤ 3:1. Popular: 1024x1024, 1536x1024,
                                1024x1536, 2048x2048, 3840x2160 (4K landscape),
                                2160x3840 (4K portrait), auto. Default: auto
--quality <quality>             low | medium | high | auto. Default: high
--model <model>                 gpt-image-2 | gpt-image-1.5 | gpt-image-1 | gpt-image-1-mini.
                                Default: gpt-image-2
--moderation <level>            auto | low. Default: auto
--n <count>                     Number of images to generate (1-4). Default: 1

Recipes (top use cases)

Three battle-tested prompt skeletons live in recipes/. Read the relevant one before composing
prompts — they encode composition rules, ratio defaults, and gotchas.

Recipe	Use when
infographic.md	Stats, frameworks, magazine-spread educational visuals
meta-ad.md	FB/IG ad creative with hook overlay + product/lifestyle context
viral-linkedin.md	High-contrast typographic scroll-stop for LinkedIn feed

Aesthetic baseline (optional)

aesthetic.md ships an editorial baseline (clean, typographic, cream/black palette).
Override or replace it for your brand. The tool does NOT auto-load aesthetic — recipes
inline the relevant style cues into the prompt.

Routing

Reach for this skill when the job is one of:

• Multi-reference composition (2+ reference images merged into one scene)
• Editorial / heavy typography (posters, magazine spreads, infographics)
• Text-accurate layouts where letter-perfect rendering matters

Gotchas

• gpt-image-2 does NOT support transparent backgrounds. Omit background: transparent.
• Edits endpoint always processes references at high fidelity → input tokens stack fast with multiple references.
• Complex prompts can take up to 2 minutes. Plan for it.
• Org verification (https://platform.openai.com/settings/organization/general) is required before first call.
• Rate limits depend on your OpenAI tier. Batch carefully.