ultimate-gemini-skill

🚀 Easiest way to install

One command — no git clone, no manual copying:

npx skills add bradAGI/ultimate-gemini-skill

That's it. Auto-detects your coding agent and drops the skill in the right place.

Works with every agent supported by skills.sh

This skill is agent-agnostic. The npx skills add installer supports 19+ coding agents / IDEs, including:

Claude Code · Cursor · Windsurf · Codex · Cline · GitHub Copilot · Gemini · VSCode · Goose · AMP · Antigravity · ClawdBot · Droid · Kilo · Kiro CLI · Nous Research · OpenCode · Roo · Trae

If your agent loads skills from a standard skills/ or similar directory, the installer will place the files correctly. You can also install manually (see below) — the skill itself is just a SKILL.md + a Python CLI, so any harness that can read a markdown skill definition can use it.

After installing, set GEMINI_API_KEY (see Securing your API key below) and run pip install google-genai pillow once for the bundled CLI.

---

A standalone, agent-agnostic skill for Gemini 3.1 Flash Image (Nano Banana 2). Ships a self-contained Python CLI — scripts/gemini_image.py — that calls the Gemini API directly. Derived from ultimate-image-gen-mcp by @anand-92 — all credit for the underlying tool design and prompt templates goes to the original author.

The banner above was generated by the skill itself — generate-image --aspect-ratio 21:9 --image-size 1K.

Samples

All three generated with this skill at 1K, rendered by scripts/gemini_image.py. Commands shown.

1. Real-world grounded photo (Google Search + Image Search)

python3 scripts/gemini_image.py generate-image \
  --prompt "Tony Hawk performing a kickflip at Venice Beach skatepark, golden hour, editorial sports photography" \
  --aspect-ratio 16:9 --image-size 1K \
  --enable-google-search --enable-image-search

The model grounded on live Google results — note the accurate "birdhouse" shirt (Tony Hawk's skate brand) and the real Venice Beach setting.

2. Logo with legible text (thinking_level: high)

python3 scripts/gemini_image.py generate-image \
  --prompt "Modern minimalist logo for a coffee brand called 'Orbit', clean circular wordmark, deep blue on white, flat vector, strong silhouette" \
  --aspect-ratio 1:1 --image-size 1K \
  --response-modalities IMAGE --thinking-level high

Gemini 3.1 Flash renders text inside images legibly — hard for most image models.

3. Cinematic 21:9 scene

python3 scripts/gemini_image.py generate-image \
  --prompt "Cinematic still of a lone astronaut walking across a bioluminescent alien beach at dusk, two moons on the horizon, ARRI Alexa anamorphic, teal and magenta color grade" \
  --aspect-ratio 21:9 --image-size 1K

What it gives your agent

• Triggers — broad description so the skill fires on "make an image", "design a logo", "give me a banner", "storyboard this", etc.
• Decision tree — single vs batch, when to enable Google Search / Image Search grounding, when to bump thinking_level: high.
• User defaults — documents overrides for image_size: 1K and aspect_ratio: 21:9 for banners.
• Parameter reference — full schema, 14 aspect ratios, sizes, formats, modalities, env vars (references/parameters.md).
• 26 prompt templates — photography, cinematic, logo, macro, fashion, storyboard, archviz, isometric, food, and more (references/templates.md).
• Bundled CLI — scripts/gemini_image.py with generate-image and batch-generate subcommands.

Subcommands

Subcommand	Purpose
generate-image	Single image
batch-generate	Up to 8 prompts, run in parallel

Flags are kebab-case (e.g. --enable-google-search, --reference-image-paths). Run with --help for the full surface.

Options at a glance

--image-size (default: 2K; user-preferred: 1K)

Value	Use for
512px	Fastest, cheapest — thumbnails, quick iteration
1K	Everyday use, good balance
2K	Server default — balanced quality
4K	Print, hero shots, final deliverables

--aspect-ratio (default: 1:1) — all 14 options: 1:1 · 2:3 · 3:2 · 3:4 · 4:3 · 4:5 · 5:4 · 9:16 · 16:9 · 21:9 · 1:4 · 1:8 · 4:1 · 8:1

• Square / social → 1:1
• Portrait → 4:5, 2:3
• Landscape → 3:2, 4:3
• Mobile / story → 9:16
• YouTube / widescreen → 16:9
• Cinematic / banner / hero → 21:9
• Pano strips → 4:1, 8:1, 1:4, 1:8

--output-format: png (default), jpeg, webp --thinking-level: minimal (default, fast), high (slower, higher quality) --response-modalities: TEXT IMAGE (default), or IMAGE only (cleaner for logos/textures) Reference images: up to 14 via --reference-image-paths (10 objects + 4 characters) Batch cap: 8 prompts per batch-generate call, run in parallel

Prerequisites

• Python 3.11+
• pip install -r requirements.txt (or pip install google-genai pillow)
• GEMINI_API_KEY (or GOOGLE_API_KEY) in the environment — free key: https://aistudio.google.com/app/apikey

Install the skill

Recommended (one command):

npx skills add bradAGI/ultimate-gemini-skill

Manual alternative (if you don't have npx or prefer to see what lands where):

git clone https://github.com/bradAGI/ultimate-gemini-skill.git
mkdir -p ~/.claude/skills/ultimate-gemini
cp -r ultimate-gemini-skill/SKILL.md \
      ultimate-gemini-skill/references \
      ultimate-gemini-skill/scripts \
      ~/.claude/skills/ultimate-gemini/

Or install the packaged dist/ultimate-gemini.skill via whatever skill installer your harness supports.

Use the CLI directly

# Single image
python3 scripts/gemini_image.py generate-image \
  --prompt "Tony Hawk kickflip" \
  --aspect-ratio 16:9 \
  --image-size 1K \
  --enable-google-search

# Parallel batch (up to 8)
python3 scripts/gemini_image.py batch-generate \
  --prompts "scene 1 ..." "scene 2 ..." "scene 3 ..." "scene 4 ..." \
  --aspect-ratio 21:9 \
  --image-size 1K

# With reference images
python3 scripts/gemini_image.py generate-image \
  --prompt "same character, new pose" \
  --reference-image-paths /abs/path/char1.png /abs/path/char2.png

Images save to $PWD/gemini_images/ by default — a subfolder of wherever you run the script, so output stays with the project you're working on. Override the location with $GEMINI_OUTPUT_DIR, the model with $GEMINI_IMAGE_MODEL, and the batch cap with $GEMINI_MAX_BATCH_SIZE.

Securing your API key

Never commit GEMINI_API_KEY. Recommended, in order:

1. Secrets manager — 1Password CLI (op read "op://Private/Gemini/api_key") or Bitwarden CLI. Key never on disk in plaintext.
2. OS keychain — secret-tool (Linux/WSL), macOS Keychain, Windows Credential Manager. Export from a ~/.bashrc wrapper.
3. Restricted file — chmod 600 ~/.secrets/gemini.env, then source it from your shell rc. Fine for single-user machines. Always add .env, *.key, and agent config files to .gitignore. Rotate immediately on exposure.

Layout

.
├── SKILL.md                 # skill triggers + workflow
├── scripts/
│   └── gemini_image.py      # standalone CLI
├── references/
│   ├── parameters.md        # full param schema
│   └── templates.md         # 26 prompt recipes
├── banner.png
└── dist/
    └── ultimate-gemini.skill

Credits

Derived from anand-92/ultimate-image-gen-mcp by @anand-92 — the original upstream project whose tool design and 26 prompt templates this skill adapts. Please star the upstream.

License

MIT — see LICENSE.