spencerpauly/awesome-cursor-skills/generating-images
Claude Code and agent skill by spencerpauly/awesome-cursor-skills
- community
Checking that the SKILL.md file is still in the repository.
Generate or edit images using the OpenAI Image API (gpt-image-2). Use when the user asks to generate, create, draw, render, illustrate, mock up, or edit an image, icon, logo, mockup, illustration, OG image, blog hero, marketing asset, or similar visual. Also use when the user supplies a reference image and asks to modify, restyle, or remix it. Triggers on: "generate an image", "create an image", "make a picture of", "edit this image", "restyle this", "make a mockup of", "draw a", "render a", "illustration of".
Install
npx vskill@latest install spencerpauly/awesome-cursor-skills/generating-imagesSKILL.md
The same SKILL.md was published earlier as spencerpauly/awesome-cursor-skills/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)
- 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.
- 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.
- Do not "fix" a missing `OPENAI_API_KEY` by reading from
.envfiles,
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:
Then export it before running the script:
Or just export it directly in your shell:
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
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
2. Image-to-image (one reference)
3. Multiple reference images
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.
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:
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.
--batchis 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
| Flag | Default | Notes | |
|---|---|---|---|
--prompt | required* | Required unless --batch is used. Always include, even when editing. | |
--image | none | Pass multiple times for multiple references. Triggers images.edit. | |
--mask | none | Optional inpainting mask (PNG with alpha). | |
--out | ./image.png | Output path; index suffix added when --n > 1. | |
--size | auto | 1024x1024, 1536x1024, 1024x1536, 2048x2048, 3840x2160, etc. Edges must be multiples of 16, max 3840px, ratio ≤ 3:1. | |
--quality | auto | low (fast drafts), medium, high (final assets). | |
--format | png | png, jpeg, webp. | |
--n | 1 | Variations of the SAME prompt in one call. | |
--batch | none | Path to JSON array of job objects; runs them in parallel. | |
--concurrency | 4 | Max parallel workers in --batch mode. |
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
lowfor quick exploration / drafts (cheapest, fastest).mediumis a good default.highonly 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-2does not support
transparent backgrounds)
- Any text that must appear, in quotes (
gpt-image-2renders 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
- Print the output path back to the user.
- Do not embed the image in markdown — Cursor displays generated files
automatically when they are written into the workspace.
- 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:
- 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
- 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.
- Brand / design tokens. If the project has a clearly defined palette,
logo, or mascot, mirror them. Quick places to check:
tailwind.config.*for brand colorsglobals.css/ theme files for CSS variablespublic/for logos / mascot assets- Any existing OG images or marketing assets
- Aspect ratio / size. Pick
--sizebased 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
This is the start of the file. Read the full SKILL.md on GitHub.
Security
- Security scanPassed, 98 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingresources/generating-images/SKILL.md
Popularity
- Repository stars
- 683
- Repository forks
- 108
- 7-day trend score
- 33
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/spencerpauly/awesome-cursor-skills/generating-images)