Nano Banana

Nano Banana — Direct Gemini API Skill

Image generation that doesn't depend on an MCP server you forgot to install. One Python script, one API key, on-disk history.

When to use

Invoke when the user asks to:

• generate an image from a text prompt;
• edit an existing image with instructions;
• refine the last image you made (the "continue editing" workflow);
• compose a multi-step series of images that need to reference each other (character consistency, before/after).

Prerequisites

1. GEMINI_API_KEY in the environment. Get one at https://aistudio.google.com/. Free tier works for prototyping; image generation is metered.
2. Python 3.9+ (uses only stdlib — no pip install needed).
3. The bundled CLI: scripts/generate.py (relative to this skill's directory).

The skill does not need an MCP server. The previous nano-banana skill required nanobanana-mcp / nano-banana-mcp via npx; this one talks to the Gemini REST API directly via urllib.

The CLI surface

scripts/generate.py <subcommand> [args]

Always pass paths absolute. The script prints the final image path to stdout — capture it and reuse.

Calling the CLI

The skill exists so you remember to call the CLI correctly. The contract:

SKILL_DIR="$(dirname "$(claude skill path nano-banana 2>/dev/null || echo ~/.claude/skills/nano-banana)")/nano-banana"
python3 "$SKILL_DIR/scripts/generate.py" generate \
  --prompt "<your prompt>" \
  --aspect 4:3 \
  --size 2K \
  --out /absolute/path/to/output.png

If you can't determine the skill's path programmatically, fall back to ~/.claude/skills/nano-banana/scripts/generate.py (user-level install) or <repo>/plugins/nano-banana/skills/nano-banana/scripts/generate.py (when running inside the marketplace plugin).

Generate

python3 .../scripts/generate.py generate \
  --prompt "Cozy coffee shop interior, watercolor illustration style, warm lighting, wooden furniture, steaming cup on table" \
  --aspect 16:9 \
  --size 2K \
  --out /tmp/coffee.png

--out is optional. The history copy is always written to ~/Documents/nanobanana_generated/<id>/image.<ext> regardless; --out just copies it where you want.

Valid aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3. Valid sizes: 1K, 2K, 4K (4K = up to 5632×3072 on gemini-3-pro-image-preview).

Edit an existing image

python3 .../scripts/generate.py edit \
  --image /path/to/source.png \
  --instructions "Transform to winter scene: add snow on ground, frost on windows, overcast sky, cool blue grading" \
  --out /tmp/winter.png

If the source image lives inside the history dir, the new entry's parent_id is set automatically so you can walk lineage later.

Continue editing the last image

# Defaults to the most recent entry (--index 0)
python3 .../scripts/generate.py continue-edit \
  --instructions "Make the lighting warmer, add golden hour glow"

# Or refine a specific past entry:
python3 .../scripts/generate.py continue-edit --index 3 --instructions "..."
python3 .../scripts/generate.py continue-edit --entry-id 1779825667-bbff19ad --instructions "..."

This is the session-aware feature the old MCP exposed as continue_editing — implemented here against the on-disk history.jsonl, so it survives session restarts.

Inspect history

python3 .../scripts/generate.py history --limit 10
# [  0] 1779825667-bbff19ad  generate  Cozy coffee shop interior, watercolor illustration ...
# [  1] 1779825612-1f2a9d04  edit      Make the lighting warmer (edit of 1779825667)

For programmatic use, add --json.

Resolve a history reference

# Path to the most recent image:
python3 .../scripts/generate.py get --index 0

# A specific meta field:
python3 .../scripts/generate.py get --index 0 --field prompt

Selecting a model

The default is gemini-3-pro-image-preview (highest quality, 4K, best text rendering). Pass --model gemini-2.0-flash-exp for the faster/cheaper Gemini 2 path. The API returns whatever resolution the model produces; aspect ratio is a hint the server may interpret or ignore depending on model version.

Prompting practice

Conceptual over prescriptive. Tell the model what and why; let it figure out how.

[Subject] + [Style] + [Mood] + [Constraints]

Aspect ratio cheat sheet:

Critical limitations

Text in images is unreliable for anything longer than 1-3 words.

Gemini's image models cannot consistently render legible multi-word text. For infographics or diagrams:

1. Keep all in-image labels short (1-3 words).
2. Leave space for any longer text and overlay it in an editor (ImageMagick, Figma, Keynote).
3. Carry the full sentence in the surrounding markdown, not the image.

Logos and brand wordmarks are even worse — never expect a generated logo to be correct. Designate space in the prompt ("clean empty space in the bottom-right for a logo") and composite the real logo file post-generation with ImageMagick.

History dir defaults

By default, all generations land under ~/Documents/nanobanana_generated/. Override with --history-dir <path> or $NANO_BANANA_HOME. Each entry is a subdirectory:

~/Documents/nanobanana_generated/
├── history.jsonl                 # append-only index, newest at bottom
└── 1779825667-bbff19ad/
    ├── image.jpg                 # or .png / .webp depending on model
    └── meta.json                 # prompt, model, parent_id, timestamps

history.jsonl is append-only; deleting individual entry dirs is fine but the JSONL line stays (and get will fail with "image_path does not exist" — clean both).

Workflows

One-shot infographic

python3 .../scripts/generate.py generate \
  --prompt "Sketchnote-style infographic explaining the 5 steps of a value stream — hand-drawn feel, single teal accent on white background, short labels only, 4:3 layout" \
  --aspect 4:3 --size 2K \
  --out /tmp/vsm-infographic.png

Iterate on lighting / composition without redrawing

# First pass
python3 .../scripts/generate.py generate --prompt "..." --out /tmp/v1.png

# Refine
python3 .../scripts/generate.py continue-edit \
  --instructions "Soften the shadows, push the accent color slightly more teal (#557373)" \
  --out /tmp/v2.png

# Refine again
python3 .../scripts/generate.py continue-edit \
  --instructions "Tighten the typography" --out /tmp/v3.png

Character consistency across multiple scenes

# Establish the character once
python3 .../scripts/generate.py generate \
  --prompt "A young woman with red curly hair, freckles, green eyes, blue jacket. Portrait, studio lighting." \
  --out /tmp/char.png

# Get the entry id you just produced
ID=$(python3 .../scripts/generate.py get --index 0 --field id)

# Subsequent scenes: edit from the established character
python3 .../scripts/generate.py continue-edit --entry-id "$ID" \
  --instructions "The same woman, now seated at a cafe reading a book, soft window light"

Calling this skill from another skill

If you're writing another skill that needs an image, don't reinvent the prompt construction — invoke this skill via the Skill tool (or shell out to the CLI directly):

python3 ~/.claude/skills/nano-banana/scripts/generate.py generate \
  --prompt "<your prompt>" --aspect <ratio> --size <size> \
  --out <where the calling skill wants the file>

That's the entire integration surface. The CLI prints the final path to stdout; the calling skill picks it up from there.

Troubleshooting

References

• API docs: https://ai.google.dev/gemini-api/docs/image-generation
• Models: gemini-3-pro-image-preview (default, 4K), gemini-2.0-flash-exp (faster), gemini-2.0-flash-preview-image-generation
• This skill's source: plugins/nano-banana/ in github.com/difflabai/marketplace

End each session by printing the path of the file you produced (or the error you hit). Don't editorialise — the caller will use the path or read the error.

Source: difflabai/marketplace — distributed by TomeVault.