Skills DirectorySeptember 2, 2026
Nano Banana Claude Skill: Which Version to Install (All Variants Compared)
Last updated: September 2, 2026
The nano banana skill lets Claude generate and edit images with Nano Banana — Google's Gemini image-generation models — straight from a conversation or a Claude Code session. We index at least 25 skills in the nano-banana family, many of them copies of each other, so "which one do I install?" is the real question. Our pick for actually generating images is NikiforovAll's nano-banana: Grade A with a perfect 100/100 security score, the most-viewed generator variant we track, maintained in a real single-author repo rather than a compilation registry, with small, auditable inline Python scripts. Install it with one command:
npx -y skills add NikiforovAll/claude-code-rules --skill nano-banana --agent claude-code
One distinction before the table: the most-installed skill in the family, saren-ai's Nano Banana (160 installs, 1,167 views), doesn't call any API at all. It's a prompt-writing guide — templates and best practices for Nano Banana. Grade A, 100/100, needs no API key. If you just want better prompts, install that one instead.
What is Nano Banana?
"Nano Banana" is the community nickname for Google's Gemini image models, and the skills in this family target three of them:
- Nano Banana — Gemini 2.5 Flash Image: fast, low-cost generation and editing
- Nano Banana Pro — Gemini 3 Pro Image: higher quality, 1K/2K/4K output
- Nano Banana 2 — Gemini 3.1 Flash Image Preview: multi-image input (up to 14 images) and Google Search grounding
A nano banana skill is a SKILL.md file that teaches Claude when and how to use one of these models — typically by running a bundled or inline Python script against the Gemini API with your own key.
All meaningful variants, compared
Here are the variants that are actually distinct, with grades from our static-analysis scanner. Each skill's page has the files, the full scan result, and the exact install command.
| Skill | Author / repo | Grade (score) | Traction | What makes it different |
|---|---|---|---|---|
| Nano Banana | saren-ai/claude-skills | A (100) | 160 installs, 1,167 views | Prompt-writing guide + template pack. No API calls, no key needed. Most-installed in the family. |
| Nano Banana | NikiforovAll/claude-code-rules | A (100) | 19 views | Our pick for generation. Inline uv Python scripts using google-genai; Gemini 2.5 Flash Image; fully auditable. |
| Nanobanana Skill | feiskyer/claude-code-settings | A (92) | 17 views, 927-star repo | Bundled Python CLI; defaults to Nano Banana Pro; 10 aspect ratios, 1K/2K/4K. One medium finding (see below). |
| Nano Banana Pro | jiangye1314/openclaw-digital-workforce | A (100) | 15 views | The standard OpenClaw-ecosystem Gemini 3 Pro Image skill — the one everybody else copies. |
| Nano Banana | nexu-io/nexu | A (100) | 3,052-star repo | Ships inside the Nexu agent framework; one skill covers Nano Banana, Pro, and 2 triggers. |
| Nano Banana | SamurAIGPT/Generative-Media-Skills | A (100) | 3,325-star repo | Routes through muapi.ai rather than Google's API directly; structured "creative brief" prompting. |
| Nano Banana / Nano Banana 2 | inference-sh/skills | A (100) | 506-star repo | Requires the inference.sh CLI. The "2" variant adds Gemini 3.1 Flash Image, up to 14 input images, and Search grounding. |
| Nano Banana | mycelium-hq/ai-brain-starter | A (96) | added Sep 1, 2026 | Newest entry; tuned for content assets — logos, thumbnails, OG images, blog headers. |
| Bex Nano Banana Pro | Demerzels-lab/elsamultiskillagent | A (100) | — | Gemini 3 Pro Image via Replicate; same repo has an OpenRouter variant and a Kling video-ad workflow. |
| Nano Banana Pro | lidge-jun/cli-jaw-skills | A (100) | — | Byte-alike copy of the standard Gemini 3 Pro Image skill above. Fine, but adds nothing. |
| Nanobanana Base | majiayu000/claude-skill-registry | B (76) | — | Re-hosted Japanese-language variant. The only non-A grade in the family — see security notes. |
About the clones: majiayu000/claude-skill-registry alone re-hosts at least six nano-banana skills, with slugs that still embed the original authors' names (dicklesworthstone, danilouchoa, charlie-818, mbanner77, termix-official). Four of them, plus lidge-jun's copy, carry the byte-identical description "Generate or edit images via Gemini 3 Pro Image (Nano Banana Pro)" — the same skill re-uploaded. We collapsed those into single rows. Install from the original author's repo, not a compilation registry: each copy is scanned as its own artifact, and as the B grade above shows, copies don't inherit the original's grade. (We excluded two false matches that only contain the word "banana", plus the PaperBanana academic-figure skills — a different tool.)
How the skill works with Claude
Once installed into .claude/skills, the skill's name and description sit in Claude's context. When you say "generate an image of…" or literally "nano banana," Claude loads the full instructions — for our pick, that means writing a short google-genai Python script, running it via uv, and saving the PNG to your project. Editing works the same way: point Claude at an existing image and describe the change. A skill is instructions, not a server — nothing to keep running.
Security notes
Every generator variant needs a Gemini API key. It lives on your machine — usually the GEMINI_API_KEY environment variable (feiskyer's variant also reads ~/.nanobanana.env) — and is sent to whatever endpoint the skill calls. That's Google for most variants; the SamurAIGPT, inference.sh, and Replicate/OpenRouter variants route prompts through third-party services instead. Specific findings:
- Nanobanana Base (B, 76/100) — flagged for modifying shell startup scripts (a critical-severity pattern: its setup writes the API key into
~/.zshrc) and destructive filesystem operations. ItsSKILL.mdeven ships a hardcoded Google API key in the setup examples. Skip it. - Nanobanana Skill (A, 92/100) — one medium finding: it installs Python packages at runtime (
pip install -r requirements.txt), which could introduce malicious dependencies. Common pattern, worth knowing about. - Everything else in the table scanned clean (A, 96–100).
Our grades come from static analysis across 36 rules and 120 patterns — full methodology on the security page.
How to install (3 steps)
- Get a Gemini API key from Google AI Studio and export it:
export GEMINI_API_KEY="..."(skip this for the saren-ai prompt guide — it needs no key). - Install the skill:
npx -y skills add NikiforovAll/claude-code-rules --skill nano-banana --agent claude-code— it lands in.claude/skillsof the current project. See installing skills for global installs. - Ask Claude for an image: "generate an image of a product mockup" — the skill triggers, runs the script, and saves the file.
More image and creative skills live in the design category.
FAQ
Is the nano banana skill free? Every skill in the table is free and open source. Generation is billed through your own Gemini API key at Google's standard pricing; Google AI Studio includes a free tier.
Which nano banana skill is best? For generating images: NikiforovAll's nano-banana (A, 100/100). For Nano Banana Pro quality with 4K and aspect-ratio control: feiskyer's nanobanana-skill, accepting its one medium finding. For prompt-writing only, no API key: saren-ai's Nano Banana.
Is the nano banana skill safe? Mostly. All but one of the meaningful variants scan Grade A; the exception is a B-grade re-hosted copy that edits shell startup files and embeds a hardcoded API key. Install from original authors, check the grade on the skill's page first, and read Are Claude Skills Safe? for the ecosystem-wide picture.
What's the difference between Nano Banana, Nano Banana Pro, and Nano Banana 2? Nicknames for successive Gemini image models: Gemini 2.5 Flash Image (fast), Gemini 3 Pro Image (higher quality, up to 4K), and Gemini 3.1 Flash Image Preview (multi-image input, Search grounding). Different skills target different models — the table's last column tells you which.