nano-banana

name	nano-banana
description	This skill should be used when the user asks to "generate an image", "create an image", "nano banana", "AI image", "make a picture", "image preset", "edit an image with AI", or mentions generating visual content using the Gemini API. Provides 25 style presets across 6 categories with configurable aspect ratios and resolutions.

Purpose

Generate and edit images via the Gemini REST API using the Python script at scripts/nano_banana.py (relative to this skill's base directory). Support 26 style presets across 6 categories (Technical, Business, Creative, UI/UX, Photography, Specialized) to produce consistently styled output without manual prompt engineering. Configure aspect ratios from 14 available options and resolution tiers ranging from 512px to 4K to match any output target.

Prerequisites

• NANO_BANANA_API_KEY environment variable set to a valid Google AI Studio API key
• Python 3.8 or higher installed
• requests library installed (pip install requests)

Quick Start

Path resolution: When this skill is loaded, Claude Code displays a "Base directory for this skill:" header. Use that absolute path to resolve script paths in all bash commands below. In the examples, $BASE_DIR represents this path.

Minimal (auto-names file using config defaults):

python3 $BASE_DIR/scripts/nano_banana.py \
  --prompt "A modern office building at sunset"

With preset and explicit output:

python3 $BASE_DIR/scripts/nano_banana.py \
  --prompt "A Victorian house with garden" \
  --output ./house_blueprint.png \
  --preset blueprint

Image editing (with reference):

python3 $BASE_DIR/scripts/nano_banana.py \
  --prompt "Convert to watercolor style" \
  --output ./watercolor_version.png \
  --preset watercolor \
  --input ./original_photo.png

Arguments Reference

Argument	Required	Default	Description
--prompt	Yes	—	Text prompt describing the image to generate
--output	No	auto	Output file path; if omitted, auto-generates from config output_dir + filename_template
--preset	No	config	Style preset slug (see Preset Reference below)
--size	No	config	Resolution tier: 512px, 1K, 2K, 4K
--aspect	No	config	Aspect ratio (see Available Aspect Ratios)
--model	No	config	Model selection: flash (fast) or pro (higher quality)
--input	No	—	Reference image path for editing (repeatable, up to 14 images)
--deck	No	—	Path to a deck spec JSON file; generates all slides (mutually exclusive with --prompt, --output, --preset)
--force	No	—	Regenerate all slides even if they already exist (only used with --deck)
--list-presets	No	—	List all available presets and exit

Default priority: CLI flag > preset default > config.json > hardcoded fallback

Preset Quick-Reference

Note: Preset defaults for aspect/size apply only when not explicitly overridden via --aspect or --size flags.

Category	Slug	Aspect	Size
Technical	blueprint	3:2	2K
Technical	exploded-view	1:1	2K
Technical	anatomical	3:4	2K
Technical	isometric	1:1	2K
Technical	system-diagram	16:9	2K
Technical	site-map	4:3	2K
Business	infographic	9:16	2K
Business	slide	16:9	2K
Business	hero-banner	16:9	2K
Business	product-mockup	1:1	2K
Business	data-viz	16:9	2K
Creative	watercolor	3:2	2K
Creative	pencil-sketch	1:1	2K
Creative	flat-vector	1:1	2K
Creative	cinematic	21:9	4K
Creative	concept-art	16:9	4K
Creative	pixel-art	1:1	1K
UI/UX	app-mockup	9:16	2K
UI/UX	icon	1:1	1K
UI/UX	logo	1:1	1K
UI/UX	wireframe	16:9	2K
Photography	portrait	3:4	2K
Photography	lifestyle	3:2	2K
Photography	architectural-viz	16:9	4K
Specialized	storyboard	16:9	2K
Specialized	sketchnotes	4:3	2K

Available Aspect Ratios

1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9

Available Resolutions

512px, 1K, 2K, 4K

Edit Mode

Pass one or more reference images using the --input flag to activate edit mode instead of pure generation:

• Pass a reference image with --input path/to/image.png; repeat the flag up to 14 times for multi-image edits
• Up to 14 reference images are supported in a single call
• Supported formats: PNG, JPEG, WebP, GIF
• Write the prompt to describe what to do with or to the reference image(s) rather than describing a scene from scratch
• Combine with presets for style transfer — for example, --preset watercolor --input photo.png applies the watercolor style system prompt alongside the reference image
• When multiple --input files are provided, the model receives all images simultaneously and can blend, compare, or composite them based on the prompt

Deck Mode

Generate an entire slide deck from a JSON specification file in a single invocation. Each slide is rendered as a separate image using per-slide prompts, presets, and presentation-level styling defined in the spec.

Invoke with --deck:

python3 $BASE_DIR/scripts/nano_banana.py --deck ./decks/my_deck_spec.json

Key behaviors:

• --deck path/to/deck_spec.json — reads the deck specification and generates all slides sequentially
• --force — regenerate all slides even if output files already exist; without this flag, existing slides are skipped (resume behavior)
• When --deck is provided, --prompt, --output, and --preset are mutually exclusive and must not be used — all settings come from the deck spec JSON and the presentation config
• --model and --size can still be passed to override the deck-level defaults

Output format:

The script creates a numbered output folder containing:

• slide_01.png, slide_02.png, ... slide_NN.png — one image per slide
• A copy of deck_spec.json — preserved alongside the output for reproducibility

Resume behavior:

If generation is interrupted, run the same command again. The script detects existing slide files (e.g., slide_03.png) and skips them automatically. Only missing slides are generated. Use --force to override this and regenerate everything.

Error handling:

• Individual slide failures do not abort the deck — the script continues to the next slide
• Failed slides are listed in the JSON summary output
• Exit code 0 if at least one slide was generated successfully
• Exit code 1 only if ALL slides failed

JSON summary output:

On completion, stdout contains a JSON summary:

{
  "deck": "/abs/path/to/output_folder",
  "slides_generated": 8,
  "slides_failed": 1,
  "failed_slides": [{"slide_number": 3, "error": "Content filtering triggered"}],
  "warnings": [{"slide_number": 5, "warning": "Generated with simplified prompt (text_panel dropped)"}],
  "total_slides": 9,
  "model": "flash",
  "presentation_config": "consulting"
}

Structured slide fields (recommended):

The deck spec supports structured fields per slide for higher-quality output:

Field	Required	Description
heading	Yes*	Bold title text at top of slide (~8 words)
visual	Yes*	Dominant physical metaphor description (~30 words)
labels	No	Array of short annotation strings (1-4 words each) placed on the visual
text_panel	No	Explanatory text in bottom panel (max 2 sentences)
reference_image	No	Path to a style reference image loaded alongside the prompt
style_overrides	No	Override aspect, size, or model per slide

*When heading and visual are present, the script uses template-based prompt assembly from the presentation config. When absent, it falls back to the legacy prompt field with prefix concatenation.

Reference images: When reference_image is set, the image is loaded and sent alongside the text prompt as a style reference. Relative paths are resolved from the current working directory. Use this to maintain visual consistency across slides or to match an external style reference.

Retry behavior: If a slide fails to generate, the script automatically retries once with a simplified prompt (text_panel removed). Successful retries are recorded in the warnings array of the JSON summary output.

Generating deck spec files:

Use the /deck-prompt skill to generate deck specification JSON files from markdown or PDF reports. That skill handles content extraction, slide planning, and per-slide prompt authoring — producing a ready-to-render spec file for --deck.

Models

Model	Flag	Best For
Gemini 3.1 Flash Image	--model flash	Fast generation, iteration, drafts
Gemini 3 Pro Image	--model pro	High-quality final assets, complex scenes

Select flash for iterative work where speed matters. Select pro for final deliverables or complex scenes with fine detail requirements.

Output

The script writes results to stdout as JSON and saves the image to the path specified by --output:

• Success: {"output": "/abs/path.png", "model": "model-id", "preset": "slug-or-null"} with exit code 0
• Error: {"error": "message", "code": "ERROR_TYPE"} with a non-zero exit code
• Image is always saved as PNG regardless of the --output extension
• Parent directories of the --output path are created automatically if they do not exist
• Version safety: If the target file already exists, the script appends _v2, _v3, etc. instead of overwriting. The JSON output reflects the actual versioned path used

Parse stdout JSON to confirm success and retrieve the resolved absolute output path.

Configuration

Edit $BASE_DIR/config.json to set persistent defaults:

{
  "output_dir": "output/nano-banana",
  "size": "2K",
  "aspect": "1:1",
  "model": "flash",
  "preset": null,
  "filename_template": "{prompt_slug}_{timestamp}.png"
}

Key	Type	Description
output_dir	string	Directory for auto-generated output (relative to project root, or absolute)
size	string	Default resolution tier (512px, 1K, 2K, 4K)
aspect	string	Default aspect ratio
model	string	Default model (flash or pro)
preset	string/null	Default preset slug (null = none)
filename_template	string	Template for auto-generated filenames. Variables: {prompt_slug}, {timestamp}

When --output is omitted, the script auto-generates a filename using output_dir + filename_template and prints the path to stderr.

Troubleshooting

Symptom	Resolution
"Missing NANO_BANANA_API_KEY"	Set the environment variable: export NANO_BANANA_API_KEY=your_key_here using a key from Google AI Studio
"Missing 'requests' library"	Install the dependency: pip install requests
Content filtering refusal	Rephrase the prompt to avoid terms that trigger Gemini safety filters; remove explicit style descriptors if necessary
Rate limiting errors	The script retries once after 2 seconds automatically; if the error persists, wait 60 seconds and retry manually
Timeout on --model pro	API timeout is 60 seconds; complex prompts with pro may approach this limit — simplify the prompt or switch to flash for drafts
Corrupt or zero-byte output	Confirm the --output path is writable and the filesystem has available space

Additional Resources

• $BASE_DIR/references/presets.md — Full prompt engineering guide with per-preset tips, example prompts, and customization advice for each of the 25 presets