Generating Images

Generating Images (OpenAI gpt-image-2)

Use this skill any time the user asks to generate or edit an image. It wraps OpenAI's gpt-image-2 model via a Python script, supports both text-only prompts and one-or-more reference images, and writes the resulting PNG/JPEG/WebP to disk.

Hard rules (do not violate)

1. Always use `gpt-image-2`. Never fall back to gpt-image-1, dall-e-3, or any other model. The script has no --model flag for this reason. 2. Fail fast on any error. Do not retry, do not swap models, do not patch around missing credentials, do not silently degrade quality. If the script exits non-zero, surface the error to the user verbatim and stop. 3. Do not "fix" a missing `OPENAI_API_KEY` by reading from .env files, 1Password, etc. unless the user explicitly tells you to. If the env var is missing, ask the user how they want to provide it (or to export it) and then stop.

When to use

• User asks for a generated image: icon, logo, illustration, mockup, OG image,

blog hero, marketing asset, concept art, diagram-style image, etc.

• User provides one or more images and asks to edit, restyle, combine, or use

them as references.

• User asks to remove/replace part of an image (use --mask).

Do not use this skill for:

• Charts/plots/data viz (generate via code instead).
• Sourcing existing photos (use a stock photo skill if available).
• Screenshots of the user's app (use a screenshot skill if available).

Prerequisites

1. OpenAI API key

You need an OPENAI_API_KEY exported in your environment. Get one at platform.openai.com/api-keys.

The skill ships with a .env.example next to this SKILL.md. Copy it and fill in your key:

cp .env.example .env
# then edit .env and put your real key in

Then export it before running the script:

set -a && source .env && set +a

Or just export it directly in your shell:

export OPENAI_API_KEY="sk-..."

If OPENAI_API_KEY is not set, the script exits with code 2 immediately. Do not try to read it from anywhere else without the user's explicit permission.

2. Org verification

Your OpenAI org must be verified for gpt-image-2 at platform.openai.com/settings/organization/general. If you see a 403 mentioning "organization must be verified", surface it and stop — do not switch models.

3. Python dependency

pip install --upgrade openai

Script location

The Python script lives next to this SKILL.md at scripts/generate_image.py. When this skill is installed at ~/.cursor/skills/generating-images/, the script will be at ~/.cursor/skills/generating-images/scripts/generate_image.py.

It prints the absolute path(s) of the written image(s) to stdout. Errors go to stderr with a non-zero exit code, and the script exits immediately on the first error.

How to invoke

Always run via the Shell tool. Pick a sensible output path inside the user's current workspace (e.g. ./public/generated/<slug>.png for web projects, or ./<slug>.png otherwise).

1. Text-to-image

python3 ~/.cursor/skills/generating-images/scripts/generate_image.py \
  --prompt "Minimal flat-vector app icon for a note-taking app, indigo gradient, rounded square, soft shadow" \
  --size 1024x1024 \
  --quality high \
  --out ./icon.png

2. Image-to-image (one reference)

python3 ~/.cursor/skills/generating-images/scripts/generate_image.py \
  --prompt "Restyle this photo as a watercolor painting with warm tones" \
  --image ./photo.jpg \
  --out ./photo-watercolor.png

3. Multiple reference images

python3 ~/.cursor/skills/generating-images/scripts/generate_image.py \
  --prompt "Photorealistic flat-lay product shot combining all of these items on a white background" \
  --image ./a.png --image ./b.png --image ./c.png \
  --out ./flatlay.png

4. Masked edit (inpainting)

The mask must be the same size and format as the first input image, with an alpha channel marking the editable region.

python3 ~/.cursor/skills/generating-images/scripts/generate_image.py \
  --prompt "Replace the sky with a vivid sunset" \
  --image ./scene.png --mask ./sky-mask.png \
  --out ./scene-sunset.png

5. Batch / parallel mode (many distinct images at once)

When you need to generate multiple different images in one go (e.g. a set of blog heroes, several icon variations with different prompts, OG images for many pages), use --batch instead of running the script N times. It runs all jobs in parallel from a single Python process — much faster than serial calls and avoids repeated SDK startup cost.

Write a JSON file describing every job, then call the script once:

cat > /tmp/img-jobs.json <<'EOF'
[
  {
    "prompt": "Minimal flat-vector app icon for a note-taking app, indigo gradient, rounded square",
    "out": "./public/icons/notes.png",
    "size": "1024x1024",
    "quality": "high"
  },
  {
    "prompt": "Photoreal blog hero: a cozy library with warm afternoon light, 5:3 ratio",
    "out": "./public/static/blog/library.png",
    "size": "1600x960",
    "quality": "medium"
  },
  {
    "prompt": "Restyle this product photo as a watercolor painting with warm tones",
    "image": ["./public/products/mug.jpg"],
    "out": "./public/products/mug-watercolor.png"
  }
]
EOF

python3 ~/.cursor/skills/generating-images/scripts/generate_image.py \
  --batch /tmp/img-jobs.json --concurrency 5

Each job object accepts the same fields as the CLI flags: prompt (required), out, size, quality, format, n, image (string or array of strings), mask. Defaults match the single-shot CLI.

Behavior:

• All jobs run concurrently up to --concurrency (default 4). A reasonable

range is 3–8; OpenAI rate-limits per org so don't go too wild.

• Each successfully written image's absolute path is printed to stdout as soon

as that job finishes, one per line.

• If any job fails, its error is printed to stderr (ERROR: job <i> failed: ...)

and the script exits with code 1 after the remaining jobs finish. Other jobs are not cancelled — partial output is fine and you can retry only the failed ones.

• --batch is mutually exclusive with --prompt / --image / --mask.

When to prefer `--batch` over parallel Shell calls: any time you're generating ≥2 distinct images in the same turn. Don't fire multiple parallel Shell invocations of this script — use one batch call instead.

Don't confuse with `--n`. --n produces multiple variations of the same prompt in a single API call (cheaper, but all the same idea). --batch runs different prompts in parallel. They can be combined: a batch job can set "n": 4 to get 4 variations of that one prompt.

Flags reference

There is intentionally no `--model` flag. The model is hardcoded to gpt-image-2.

Sizing guidance

• App icons / square thumbnails → 1024x1024
• Landing-page heroes / OG images → 1536x1024
• Blog hero (5:3) → 1600x960 (both edges multiples of 16, ratio = 5:3)
• Mobile / portrait illustrations → 1024x1536
• Marketing posters / 4K assets → 3840x2160

Quality guidance

• low for quick exploration / drafts (cheapest, fastest).
• medium is a good default.
• high only for final, ship-ready assets — significantly more expensive

and can take up to ~2 minutes.

If the user just says "generate an image" with no signal of finality, default to --quality medium.

Prompt-writing tips

For best results, include in the prompt:

• Subject (what is in the image)
• Style (flat vector, watercolor, photoreal, isometric, line drawing, 3D render…)
• Composition / camera (close-up, top-down, wide shot)
• Color palette / mood
• Background (white, gradient, scene — note: gpt-image-2 does not support

transparent backgrounds)

• Any text that must appear, in quotes (gpt-image-2 renders text well)

If the user gives a vague prompt, expand it with sensible defaults rather than asking back, unless the request is genuinely ambiguous.

After generating

1. Print the output path back to the user. 2. Do not embed the image in markdown — Cursor displays generated files automatically when they are written into the workspace. 3. If the result is meant for a website/app, consider also running it through an optimizer (e.g. pngquant, cwebp) when file size matters.

Gather context BEFORE generating

Unless the user has spelled out exactly what they want (subject, style, palette, size, destination), do a quick context-gathering pass first. The goal is for the generated image to feel like it belongs where it's going, not like a random asset dropped into the project. Skipping this step is the #1 way this skill produces off-brand results.

Things to look at, in roughly this order:

1. Sibling images at the destination. If the image will live in public/static/blog/, public/static/marketing/, assets/, etc., open one or two existing images in that folder with the Read tool. Match their:

• Illustration style (3D cartoon, flat vector, photoreal, line art, isometric…)
• Color palette and lighting
• Subject conventions (e.g. "always features the product mascot", "always a

metaphor, never literal screenshots", etc.)

• Aspect ratio and resolution

2. The surface that will display it. Read the relevant file:

• Blog post → read the MDX/Markdown (title, tags, opening paragraphs, key metaphors).
• Landing page section → read the component, headline, and surrounding copy.
• README → read the top of the README.
• Component → read the component to understand what it represents.

Pull the image's meaning from the actual content, not just the filename.

3. Brand / design tokens. If the project has a clearly defined palette, logo, or mascot, mirror them. Quick places to check:

• tailwind.config.* for brand colors
• globals.css / theme files for CSS variables
• public/ for logos / mascot assets
• Any existing OG images or marketing assets

4. Aspect ratio / size. Pick --size based on the surface: blog hero, OG image, square avatar, mobile portrait, etc. Match what's already there.

Then write the prompt incorporating what you learned: subject pulled from the content, style + palette pulled from sibling assets and brand tokens, composition matched to the surface.

If the user did give explicit direction (style, colors, exact subject), honor it and skip context-gathering. If they gave partial direction, gather context for the parts they left open.

Don't ask the user clarifying questions for things you can reasonably infer from the codebase — infer first, ask only when something is genuinely ambiguous (e.g. two equally valid styles already exist in the project).

Place it AND wire it up — don't just dump a file

When the user asks for an image for a specific surface (a blog post, a landing page, an OG card, a README, a component, etc.), you are responsible for the whole job, not just the PNG. Always do these in order:

1. Pick the correct on-disk location for that surface. Look at what already exists and match it. Examples:

• Blog hero → wherever existing blog images live (e.g.

apps/<app>/public/static/blog/<slug>.png).

• Landing page asset → wherever other landing assets live (e.g.

apps/<app>/public/static/marketing/...).

• README / docs image → docs/images/, assets/, or next to the doc.
• Component-specific asset → next to the component or in its

public//assets/ folder.

Use the file's slug, component name, or section name for the filename. Don't invent a new convention if one already exists.

2. Wire the image up so it actually shows where the user wanted it. This is not optional. Examples:

• Blog post MDX → update the image: (or equivalent) frontmatter field to

point at the new path. Replace any placeholder Unsplash/stock URL.

• Landing page section → import or reference the new asset in the relevant

component/JSX.

• OG image → update the <meta property="og:image"> / metadata config.
• README → add the appropriate Markdown image tag.

3. Match existing conventions for paths (relative vs /static/... vs @/assets/...), file format (png/webp/jpg), and any wrapper components (next/image, custom <Image>, etc.).

4. Don't ask first. If the user asked for an image for a known surface, do the placement + wiring automatically and tell them what you changed at the end. Only ask when the destination is genuinely ambiguous.

Errors — surface, don't hide

If any of the following happen, stop immediately and report the error to the user. Do not retry, do not change the model, do not change the prompt.

• OPENAI_API_KEY is not set → ask the user how to provide it.
• openai package not installed → tell the user to run pip install --upgrade openai.
• 403 "organization must be verified" → tell the user to verify at

platform.openai.com/settings/organization/general. Do not switch models.

• 400 size error → report it; let the user pick a valid size.
• 400 about transparent background → report it; gpt-image-2 doesn't

support transparency.

• Any other API error → report verbatim and stop.