| name | generate-image |
| description | This skill should be used when the user asks to "generate an image", "create a banner", "make artwork", "create an illustration", "generate a logo", "make a graphic", "design a header", "AI art", "img2img", "social share image", "OG image", "open graph image", "Twitter card", "hero image", "cover photo", "profile picture", "social media graphic", or needs AI image generation. Handles prompt rewriting and Gemini 3 Pro image generation API calls. |
Generate Image
Generate images using Nano Banana Pro (gemini-3-pro-image).
When to Use
Use this skill when the user asks to:
- Generate an image from a text prompt
- Create artwork, illustrations, or graphics
- Generate variations of an existing image (img2img)
- Create scenes with multiple reference images (character + location + other characters)
Style Selection
If the user hasn't specified a style, present a multi-choice question before proceeding:
How would you like to handle the art style?
- Pick a style - Browse 167 styles visually and choose one
- Let me choose - I'll suggest a style based on your prompt
- What are styles? - Learn how the style system works
- No style - Generate without a specific art style
Use the AskUserQuestion tool to present this choice.
If the user picks "Pick a style", launch the style picker (pick mode is the default):
STYLE_JSON=$(bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/browsing-styles/scripts/preview_server.ts)
The picker opens a browser. The user clicks a style, views the detail modal, clicks "Select This Style". STYLE_JSON receives the selection as JSON and the server exits. Parse the id field and pass it via --style <id> to the generate command.
If the user picks "Let me choose", select the most fitting style based on their prompt and explain why. Use the style suggestions in the "Suggest Styles to Users" section below.
If the user picks "What are styles?", explain: there are 167 curated art styles across 10 categories (traditional, digital, illustration, photography, design, retro, technique, decorative, creative, cultural). Each style includes prompt hints and a visual reference tile that gets sent to Gemini for better adherence. Then re-ask the style question.
If the user already specified a style (e.g. "generate in watercolor style"), skip this step entirely and use --style directly.
Gather Design Direction First
Ask clarifying questions before rewriting prompts to understand the user's intent:
- Purpose: What is the image for? (banner, logo, social media, product shot, art piece)
- Color palette: Any brand colors? Dark/light theme? Specific mood?
- Composition constraints: Aspect ratio needs? Text overlay space? Full-bleed vs bordered?
- Key elements: What must be included? What should be avoided?
Simple requests like "make a cat image" can proceed with sensible defaults. Complex requests like "create a banner for my app" require clarification to avoid iteration waste.
Prompt Rewriting (Critical)
Before generating any image, always rewrite the user's prompt using the guide in references/prompt-guide.md.
The core principle: "Describe the scene, don't just list keywords."
Rewriting Checklist
Transform simple prompts by adding:
- Subject details: Specific appearance, clothing, expression, materials
- Environment/Setting: Location, time of day, weather, indoor/outdoor
- Lighting: Natural/artificial, direction, quality, color temperature
- Composition: Camera angle, distance, framing, focal length
- Style/Aesthetic: Photorealistic, illustration style, art movement
- Mood/Atmosphere: Emotional tone, color palette mood
- Technical specs: Aspect ratio, level of detail, textures
Example Transformation
User says: "banner for my app"
Rewritten prompt: "A modern, full-bleed banner for a technology application. Dark gradient background transitioning from deep navy to black. Abstract geometric network visualization with glowing nodes connected by thin lines. Clean sans-serif typography positioned left of center. Professional, tech-forward aesthetic with no visible borders or edges - the design extends seamlessly to all edges. 16:9 aspect ratio."
User says: "a cat"
Rewritten prompt: "A fluffy orange tabby cat lounging on a sun-drenched windowsill, soft afternoon light creating a warm glow on its fur. The cat is in a relaxed pose with half-closed eyes, conveying contentment. Shot from a low angle with shallow depth of field, the background showing a blurred garden view. Photorealistic, warm color palette."
Recommended Workflow
Draft → Iterate → Final approach saves time and API costs:
- Draft Phase (1K): Generate quickly at default resolution to test prompts
- Iteration Phase: Refine prompts incrementally, creating new files each attempt
- Final Phase (4K): Only produce high-res output after prompt is validated
Do not read the image back - the script outputs only the file path. This allows efficient iteration without Claude loading large image data.
Usage
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "prompt" [options]
Options
--input <path> - Reference image (can specify multiple times, up to 14 images)--style <id> - Apply style from the style library (see browsing-styles skill)--size <1K|2K|4K> - Image size (default: 1K for fast drafts)--aspect <ratio> - Aspect ratio: 1:1, 16:9, 9:16, 4:3, 3:4--negative <prompt> - Negative prompt (what to avoid)--count <n> - Number of images (1-4, default: 1)--guidance <n> - Guidance scale--seed <n> - Random seed for reproducibility--output <path> - Output path--provider <name> - gemini, openai, or xai. Omit to auto-pick the best available provider for the request (see Models). --model is a legacy alias; grok maps to xai.
Examples
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "cyberpunk cityscape at night"
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "mountain landscape" --style impressionism
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "portrait" --style ukiy
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "city street" --style noir
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "mountain landscape" --size 4K --aspect 16:9
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "portrait of a cat" --negative "low quality, blurry"
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "cat sleeping" --style wtrc --size 4K --aspect 1:1
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "abstract art" --count 4
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "make it look like a watercolor painting" --input photo.jpg
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "King on throne in dramatic lighting" \
--input character.png \
--input throne-room.png \
--aspect 16:9 --size 2K
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "Two warriors facing each other in combat" \
--input warrior1.png \
--input warrior2.png \
--input battlefield.png \
--aspect 16:9
Multiple Reference Images
Gemini supports up to 14 reference images per request:
- 6 objects - Locations, items, environments
- 5 humans - Character consistency across generations
- Use detailed prompts describing how reference images should be combined
- Reference images help maintain consistency across scene generations
Common Formats
| Format | Aspect Ratio | Recommended Size | Crop Target |
|---|
| OG / Social Share | 16:9 | 2K | 1200x630 via sips |
| Twitter Card | 16:9 | 2K | 1200x628 via sips |
| Hero Banner | 16:9 | 4K | Full width |
| Profile Picture | 1:1 | 1K | 512x512 or 1024x1024 |
| Story / Reel | 9:16 | 2K | 1080x1920 |
| Team Lineup | 21:9 | 2K | Full width |
Social Share Image Workflow
All 4 steps are required. Do not skip optimization.
- Generate at 16:9 with center-weighted prompt at
--size 2K (retina-ready):
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "CENTER all important elements, no content near edges. [your prompt]" --aspect 16:9 --size 2K --output social-share.png
- Crop to exact platform dimensions:
sips -z 630 1200 -c 630 1200 social-share.png --out social-share-og.png
sips -z 628 1200 -c 628 1200 social-share.png --out social-share-tw.png
- Optimize - convert to JPEG at 85% quality (social images never need transparency):
sips -s format jpeg -s formatOptions 85 social-share-og.png --out social-share-og.jpg
- Verify dimensions and file size:
sips -g pixelWidth -g pixelHeight social-share-og.jpg
ls -la social-share-og.jpg
Why optimize? Gemini outputs PNG (~1-2MB). JPEG at 85% reduces to ~200-400KB with no visible quality loss. Social platforms re-compress uploads, so starting with a well-optimized JPEG prevents double-compression artifacts.
Available Styles
100+ styles across 9 categories: traditional art, digital, illustration, photography, design/UI, retro, techniques, decorative, and creative/material-based.
When a style has a tile reference image (assets/tiles/<id>.png), it is automatically sent as a visual reference to Gemini alongside the prompt hints — this produces much better style adherence.
Suggest Styles to Users
When generating images, proactively suggest relevant styles:
- For landscapes:
impr, roma, wtrl, cine
- For portraits:
barq, prer, anim, fnoi
- For UI/web:
brut, nbrt, glas, flat
- For retro/vintage:
y2k, frug, psyc, koda
- For whimsical/fun:
ghbl, kawa, chld, brit
- For dramatic:
cybr, nnoi, surr, expr
- For creative:
sand, undr, orig, clay, lego
Browse all: bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/browsing-styles/scripts/list_styles.ts --table
Search: bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/browsing-styles/scripts/list_styles.ts --search "watercolor"
Models & Providers
Three providers, selected with --provider or auto-picked when omitted.
| Provider | Model | Key | Strengths | Can't do |
|---|
gemini (default fallback) | gemini-3-pro-image (Nano Banana Pro) | GEMINI_API_KEY | Style tiles, up to 14 reference images, negative prompts, transparency, 1K/2K/4K | — |
openai | gpt-image-2 | OPENAI_API_KEY | Best in-image text, dense prompt understanding, custom sizes, image-to-image (--input, up to 16, via the edits endpoint) | No transparency, no style tiles, no negative param |
xai | grok-imagine-image-quality | XAI_API_KEY | Fast, spicier; good when Gemini's filter blocks a benign prompt | Text-to-image only — no img2img/tiles/negative |
Auto-pick (default when --provider is omitted)
The script chooses the best provider whose key is present and that supports
what the request needs. Ranking is openai > gemini > xai.
--input (img2img / reference images) is supported by both openai and gemini
— openai routes it through the gpt-image-2 edits endpoint, gemini through Nano
Banana Pro (which also supports style tiles + up to 14 refs). So img2img
auto-picks openai when its key is present; pass --provider gemini for Gemini's
style-tile / multi-ref compositing.
Capability gating still forces Gemini when the request uses --style (style
tiles) or --negative (a true negative parameter) — only Gemini supports those.
Pin a default with /gemskills:setup or GEMSKILLS_IMAGE_PROVIDER.
When openai is chosen with a --style, the style tile image is dropped (the
textual hints are still applied) and --negative is folded into the prompt as an
Avoid: … clause. xAI additionally drops reference images (text-to-image only).
Prompt templates (tune per provider)
Each provider parses prompts differently. After the provider is resolved, read
the matching guide and rewrite the prompt to it:
providers/prompts/image.gemini.mdproviders/prompts/image.openai.mdproviders/prompts/image.xai.md
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "neon city street, rain" --provider openai --size 4K
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "watercolor fox" --provider xai
Models verified live: June 2026 (gpt-image-2, grok-imagine-image-quality, gemini-3-pro-image). If a newer generation exists, STOP and suggest a PR to b-open-io/gemskills.
Reference Files
For detailed prompting strategies and techniques:
references/prompt-guide.md - Comprehensive Gemini prompting guide with 7 strategies, example transformations, and best practices from Google's official documentation- ask-gemini skill's
references/gemini-api.md - Current Gemini models, SDK info, and dynamic documentation via llms.txt
