Nanobanana

Nano Banana Image Generation

Generate images via the Gemini CLI's nanobanana extension. Claude has no native image-rendering ability, so this CLI is the only way to actually produce an image file — don't fall back to describing an image in prose when the user wants a real asset.

When to Use This Skill

Use it whenever the user wants an actual rendered image: an illustration, thumbnail, featured image, banner, icon, favicon, rendered diagram, pattern/texture, restored photo, or story/comic panels. Trigger on verbs like generate, create, make, draw, design, visualize, restore, or edit when the target is an image.

Skip it when the user wants something that only looks adjacent:

• Code/markup diagrams (Mermaid, Graphviz, ASCII art, hand-written SVG) — those are text.
• Charts/graphs plotted from a dataset (use a plotting library, not image generation).
• UI components or design systems in code.

Before First Use

1. Confirm the extension is installed:

   gemini extensions list | grep nanobanana
   

2. If missing, install it:

   gemini extensions install https://github.com/gemini-cli-extensions/nanobanana
   

3. API key: the extension stores its key either in the GEMINI_API_KEY env var or in the OS keychain (configured via the extension settings). Don't gate on $GEMINI_API_KEY being set — a keychain-stored key won't show up there but still works. Only act if a real run fails with an auth error, in which case set the key:

   export GEMINI_API_KEY="your-key"
   

The --yolo flag

Always pass --yolo. The nanobanana commands call MCP tools that otherwise pause for an interactive approval prompt; since this skill runs non-interactively, that prompt would hang the run. --yolo auto-approves the tool calls so generation completes unattended.

Command Selection

Commands and their real options

These option lists are validated by the extension's command parser — passing an option or value outside the allowed set makes the command return an error instead of an image, so stick to exactly what's listed here.

These values are pinned to nanobanana v1.0.12. If a command returns an "Invalid option(s)" error for something documented here, the extension has likely been updated and its option set has drifted — trust the command's own error output (or gemini "/<command> --help") over this table, and adjust the call accordingly.

/generate — text-to-image

gemini --yolo "/generate 'prompt'"

• --count=N — N variations, 1-8 (default 1)
• --styles="..." — comma-separated, from: photorealistic, watercolor, oil-painting, sketch, pixel-art, anime, vintage, modern, abstract, minimalist
• --variations="..." — comma-separated, from: lighting, angle, color-palette, composition, mood, season, time-of-day
• --format=grid|separate (default separate)
• --seed=123
• --preview

There is no --aspect or --size option for /generate. To influence aspect ratio or framing, describe it in the prompt itself (e.g. "wide 16:9 banner composition", "square thumbnail", "tall vertical 9:16 poster").

/icon — app icons, favicons, UI elements

gemini --yolo "/icon 'minimalist productivity app logo' --sizes='64,128,256,512' --type='app-icon' --corners='rounded'"

• --sizes="..." — from: 16, 32, 64, 128, 256, 512, 1024
• --type="app-icon|favicon|ui-element" (default app-icon)
• --style="flat|skeuomorphic|minimal|modern" (default modern)
• --format="png|jpeg" (default png)
• --background="transparent|white|black" or a color name (default transparent)
• --corners="rounded|sharp" (default rounded)
• --preview

/diagram — rendered technical diagrams

gemini --yolo "/diagram 'user authentication flow with OAuth' --type='flowchart' --style='professional'"

• --type="flowchart|architecture|network|database|wireframe|mindmap|sequence" (default flowchart)
• --style="professional|clean|hand-drawn|technical" (default professional)
• --layout="horizontal|vertical|hierarchical|circular" (default hierarchical)
• --complexity="simple|detailed|comprehensive" (default detailed)
• --colors="mono|accent|categorical" (default accent)
• --annotations="minimal|detailed" (default detailed)
• --preview

(Note: --style here does not accept modern — that's an /icon value.)

/pattern — seamless textures and patterns

gemini --yolo "/pattern 'subtle geometric tech background' --size='512x512' --type='seamless'"

• --size="WxH" (e.g. 128x128, 256x256, 512x512; default 256x256)
• --type="seamless|texture|wallpaper" (default seamless)
• --style="geometric|organic|abstract|floral|tech" (default abstract)
• --density="sparse|medium|dense" (default medium)
• --colors="mono|duotone|colorful" (default colorful)
• --repeat="tile|mirror" (default tile)
• --preview

/story — sequential / narrative panels

gemini --yolo "/story 'a seed growing into a tree' --steps=4 --layout='comic'"

• --steps=N — 2-8 (default 4)
• --type="story|process|tutorial|timeline" (default story)
• --style="consistent|evolving" (default consistent)
• --layout="separate|grid|comic" (default separate)
• --transition="smooth|dramatic|fade" (default smooth)
• --format="storyboard|individual" (default individual)
• --preview

/edit and /restore — modify an existing image

gemini --yolo "/edit file.png 'remove the background'"
gemini --yolo "/restore old_photo.jpg 'fix scratches and restore color'"

Both take a filename, then an instruction in quotes, then optionally --preview. They accept no other options.

Sizing / aspect ratio

Only some commands take explicit dimensions:

• /icon → --sizes
• /pattern → --size=WxH

/generate, /diagram, and /story have no dimension flag — convey the desired shape/ratio in the prompt text instead (e.g. "wide 16:9 banner", "square 1:1", "vertical 9:16 story format", "1200x630 social-preview composition").

Model Selection

Default model: gemini-3.1-flash-image-preview.

To switch models, set NANOBANANA_MODEL:

# Higher quality / better reasoning (Nano Banana Pro)
export NANOBANANA_MODEL=gemini-3-pro-image-preview
# Older v1 flash model
export NANOBANANA_MODEL=gemini-2.5-flash-image

Output Location

Generated images are saved to ./nanobanana-output/ (created automatically) relative to the current working directory. After a run, list that directory, report the absolute path of the new file(s) to the user, and present the most recent one.

Examples

# Blog featured image — modern illustration
gemini --yolo "/generate 'modern flat illustration of a developer coding at a laptop, purple and blue gradient background, minimalist, wide 16:9 banner composition, no text' --preview"

# Three editorial-photo variations
gemini --yolo "/generate 'editorial photo of a coffee cup beside a laptop on a wooden desk, morning sunlight, shallow depth of field, no text' --count=3 --styles='photorealistic'"

# App icon set
gemini --yolo "/icon 'minimalist app logo for a productivity tool' --sizes='64,128,256,512' --type='app-icon' --corners='rounded'"

# Auth flow diagram
gemini --yolo "/diagram 'user authentication flow with OAuth' --type='flowchart' --style='professional'"

Refinements and Iterations

• "Try again" / "give me options" → regenerate with --count=3.
• "Make it more [adjective]" → adjust the prompt text and regenerate.
• "Edit this one" → gemini --yolo "/edit nanobanana-output/<file>.png 'adjustment'".
• "Different style" → set --styles="..." from the allowed list for that command.

Prompt Tips

1. Be specific about style, mood, colors, and composition.
2. Add "no text" when you don't want rendered words in the image.
3. Reference concrete styles ("editorial photography", "flat illustration", "3D render", "watercolor").
4. State the intended shape/ratio in words, since /generate has no aspect flag.

Troubleshooting