How to use the Nano Banana Pro Image Generator agent skill

What does the nano-banana-pro skill make the agent do?

Once installed, the agent can generate a new image from a text description, or edit an existing image you point it to. Both tasks go through Google's Nano Banana 2 API (Gemini 3.1 Flash Image). The skill runs a bundled Python script that calls this API, saves a PNG to your current folder, and reports the full file path.

The skill also follows a draft-then-final workflow: it makes quick 1K drafts while you refine the prompt, then produces a 4K final once the prompt is settled.

• Generate new images from a text description
• Edit an existing image using written instructions
• Save output as PNG in the folder you are working in
• Iterate with fast 1K drafts before a 4K final

How do you install the nano-banana-pro skill?

Download the ZIP of the skill from this page. Then add it to one of the following locations, depending on how you use your agent:

For Claude Code, unzip the folder into ~/.claude/skills/ to make it available everywhere, or into a project's .claude/skills/ folder to limit it to that project.

In the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal (all projects): ~/.claude/skills/
• Single project: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills (upload)

What do you need before it works?

The skill needs four things: uv (a Python package runner), Python 3.10 or newer, a Google Gemini API key, and internet access to Google's Gemini API. uv and the Python libraries google-genai and Pillow are installed automatically when the script runs through uv.

Your API key can be set as the GEMINI_API_KEY environment variable or passed with the --api-key option. Before running, confirm uv is available with the check below.

• uv (Python package runner)
• Python 3.10 or newer
• google-genai and Pillow (installed by uv)
• Google Gemini API key (GEMINI_API_KEY or --api-key)
• Internet access to Google's Gemini API

How do you trigger the skill with a natural request?

You do not need to type commands. Ask the agent in plain language, for example to create an image of a subject, or to edit a photo you have named. The agent chooses the script call, the filename and the resolution.

For editing, name the image file or refer to an image in the current directory. Describe the change you want, such as making the sky more dramatic or changing the photo to a cartoon style. The skill works best when you describe one change at a time.

• "Create an image of a serene Japanese garden with cherry blossoms, high resolution."
• "Edit original-photo.jpg: make the sky more dramatic with storm clouds."
• "Make a quick draft of a robot mascot first."

How do output sizes and resolutions work?

The skill supports four output sizes. If you do not mention a resolution, it uses 1K. Phrases such as thumbnail or tiny map to 512, while phrases such as high-res or ultra map to 4K. The agent follows this mapping when it builds the script call.

• 512: compact, about 512px class (for thumbnail or tiny requests)
• 1K (default): about 1024px
• 2K: about 2048px
• 4K: about 4096px (for final, high-resolution output)

What does a generate or edit command look like?

The skill runs the bundled script through uv, using the absolute path to the file. Run it from your working folder so the image is saved where you are working, not inside the skill folder. The first command generates a new image, and the second edits an existing one.

• Use an absolute path to the script; do not cd into the skill folder first.
• Always give a new filename; existing files are rejected.
• Run from your current directory so the output lands there.

What safety and privacy points should you know?

This skill runs a script and uses the network. Every request is sent to Google's Gemini API, which is a paid remote service, so the prompt and any input image leave your machine. Do not use it on sensitive or private photos you do not want sent to a third-party cloud service.

The script rejects existing files, hard links and symlinks as output targets, and it rejects symlinked path components. It creates parent directories as needed. Absolute paths and paths outside the current directory work on macOS and Linux. The skill's own offline tests can be run with Python and Pillow installed.

• Image and prompt data are sent to Google's Gemini API.
• Requests are billed to your Gemini account and can hit quota limits.
• Output will not overwrite existing files.
• Review the script before running it if you want to confirm its behavior.