gbalabanli/nano-banana

Nano Banana

Use this skill for image generation work through the Gemini CLI Nano Banana extension.

Workflow

1. Confirm the request shape.

• Identify whether the user wants generate, edit, restore, icon, diagram, pattern, or story.
• Extract the asset goal, target aspect ratio, style direction, and number of variations.
• If the user gives only a rough idea, turn it into one precise prompt plus one fallback variation.

2. Verify the local toolchain before first use.

• Read setup.md if the extension has not been verified in this environment.
• Confirm gemini is installed, the nanobanana extension is available, and GEMINI_API_KEY is set.
• If setup is missing, stop and report the exact missing prerequisite.

3. Generate or edit with one clear command.

• Prefer the most specific slash command for the job:
  • /generate for text-to-image
  • /edit for modifying an existing image
  • /restore for repair and cleanup
  • /icon for app or UI icon work
  • /diagram for flows, architecture, and structured graphics
  • /pattern for seamless backgrounds and textures
  • /story for multi-frame or narrative output
• Use gemini --yolo so the command can run non-interactively from Codex or OpenCode.

4. Present strong first-pass prompts.

• Keep prompts concrete: subject, composition, style, lighting, palette, and exclusions.
• Add no text when text rendering is not desired.
• Use prompt-patterns.md for prompt structures by output type.

5. Iterate deliberately.

• When the user asks for changes, modify only the dimensions that changed: composition, style, palette, subject detail, or output count.
• For try again or give me options, regenerate with --count=3.
• For direct revisions to an existing file, prefer /edit over a fresh /generate.

6. Return the result clearly.

• Report the command used in concise form.
• Point to the output directory or generated file path.
• Note the most important tradeoff or assumption if the prompt required interpretation.

Output Rules

• Do not claim Nano Banana is available until the local CLI and extension are verified.
• Do not route image work to another generator when this skill is explicitly requested.
• Prefer one production-quality prompt over a long brainstorm unless the user asked for options.
• Keep prompts editable and reusable; avoid overly poetic phrasing that weakens control.
• If the request is ambiguous, make one reasonable assumption and state it.

References

• Setup and verification: setup.md
• Prompt structures and examples: prompt-patterns.md