菜单
Generate and edit images using Gemini's image generation API (Imagen 3). This skill should be used when creating images, illustrations, diagrams, editing existing images, or iteratively refining visual content through multi-turn conversations.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | nano-banana |
| description | Generate and edit images using Gemini's image generation API (Imagen 3). This skill should be used when creating images, illustrations, diagrams, editing existing images, or iteratively refining visual content through multi-turn conversations. |
| allowed-tools | Bash(python:*), Read, Write |
This skill provides comprehensive access to Gemini's image generation capabilities, including text-to-image generation, image editing, multi-turn refinement, batch processing, and high-resolution output.
GOOGLE_API_KEY or GEMINI_API_KEY environment variable.venv/) with google-genai pre-installed~/.claude/skills/nano-banana/scripts/generate_image.py| Model | ID | Best For |
|---|---|---|
| Flash | gemini-2.5-flash-image | Fast generation, high volume, low latency |
| Pro | gemini-3-pro-image-preview | Professional quality, 4K, advanced reasoning |
| Capability | Script | Description |
|---|---|---|
| Image Generation | generate_image.py | Create images from text prompts |
| Image Editing | edit_image.py | Modify existing images with instructions |
| Batch Generation | batch_generate.py | Generate multiple images in parallel |
| Multi-turn Editing | chat_session.py | Iterative refinement via conversation |
python scripts/generate_image.py "a sunset over mountains" sunset.png
With options:
python scripts/generate_image.py "modern office" office.png --model pro --aspect 16:9 --size 4K
python scripts/edit_image.py photo.png "remove the background" result.png
Create a prompts file (prompts.json):
[
{"prompt": "slide 1: intro graphic", "filename": "slide_01.png", "aspect": "16:9"},
{"prompt": "slide 2: data visualization", "filename": "slide_02.png", "aspect": "16:9"}
]
Generate all:
python scripts/batch_generate.py prompts.json ./output/ --parallel 3
Or use a simple text file (one prompt per line):
python scripts/batch_generate.py prompts.txt ./output/ --aspect 16:9
Start an interactive session:
python scripts/chat_session.py --output-dir ./images/
Resume a previous session:
python scripts/chat_session.py --session-file session.json
Send a single refinement:
python scripts/chat_session.py --session-file session.json --message "make it more vibrant"
Generate a single image from a text prompt.
Usage: python scripts/generate_image.py "prompt" output.png [options]
Options:
--model, -m Model: flash (default) or pro
--aspect, -a Aspect ratio: 1:1 (default), 16:9, 9:16, 21:9, 4:3, 3:4
--size, -s Resolution (pro only): 1K, 2K, 4K
Edit an existing image using text instructions.
Usage: python scripts/edit_image.py input.png "instructions" output.png [options]
Options:
--model, -m Model: flash (default) or pro
Generate multiple images from a prompts file.
Usage: python scripts/batch_generate.py prompts.json output_dir/ [options]
Options:
--model, -m Model: flash (default) or pro
--aspect, -a Default aspect ratio: 1:1 (default)
--parallel, -p Number of parallel workers: 1 (default)
--json Output results as JSON
Multi-turn image generation/editing session.
Usage: python scripts/chat_session.py [options]
Options:
--model, -m Model: flash (default) or pro
--session-file, -s Path to session state file (JSON)
--output-dir, -o Directory for output images: . (default)
--initial, -i Initial prompt to start with
--message Send single message (non-interactive)
flash model for quick iterations and high volume.pro model for final production assets and 4K output.--parallel for multiple images.For detailed API documentation, see references/gemini-api.md. For prompting techniques, see references/prompt-guide.md. For common editing patterns, see references/editing-patterns.md.
Final adversarial audit of finished work by one top-tier model (Fable 5 or Opus 4.8), in a single report-only pass across three axes: verify the load-bearing invariants (especially any you could NOT run locally), simplify without losing properties, and hunt for live variants of the bug shapes just fixed. You then verify each finding and act. Use as the closing gate on a substantive or risky change before calling it done, especially when part of it (a systest, a migration, a concurrency path) can't be validated locally. Distinct from /advisor (decision consult), /review-loop (multi-agent fix loop), and /variant-analysis (single-pattern sweep). Invoke via /advisor-review or after finishing hard-to-verify work.
Line-level git staging and non-interactive rebase for agents that know exactly which lines they changed. Use when you need atomic commits carved out of a larger diff, or when you need to squash/drop/reorder/reword commits without interactive prompts.
On-demand consultation of a higher-tier model (Opus 4.8 or Fable 5) from a cheaper executor session. Emulates the native advisor-tool pattern in Claude Code: keep the main loop on a cheap model (Sonnet 5) and call the expensive model rarely, via a focused consultation packet, for a plan/decision/'what am I missing' — not for code. Use when the executor is stuck, before committing to a plan, after repeated failed attempts at the same fix, or before a risky/irreversible change. Invoked via /advisor or when the executor recognizes a consult checkpoint.
Expensive planner (Opus 4.8 / Fable 5) decomposes a task into independent work items, a fleet of cheap Sonnet 5 (or Haiku 4.5) workers executes them in parallel via a Workflow, and the planner synthesizes the results. Emulates the native orchestrator pattern in Claude Code: keep the expensive model scarce (one planning pass, one synthesis pass) and the cheap model abundant (N parallel workers). Prefers a deterministic dynamic workflow when available; falls back to in-instance Task dispatch. Use when the user types /orchestrate or asks to decompose, fan out, or parallelize a task across independent workers.
Babysit CI after a push: a monitor→classify→remediate state machine that watches a run to completion, mechanically fixes trivial/build breakage, reproduces test failures locally to tell a flake from a real bug, reruns suspected flakes within a budget, hands reproduced failures to a fixer that commits a fixup and pushes, and loops until CI is green or it can justify a bail back to the user. Prefers a deterministic dynamic workflow when available; falls back to in-instance Task dispatch. Use when the user types /ci-loop or asks to watch/babysit/auto-fix CI for a pushed branch or PR until it's green.
Adversarial review→triage→fix loop until a cold verifier signs off. Fans out lens-specific reviewer subagents, verifies every finding against the code (killing false positives), auto-applies confirmed fixes as fixup commits, and repeats until a fresh verifier approves. Prefers a deterministic dynamic workflow when available; falls back to in-instance Task dispatch. Use when the user types /review-loop or asks to adversarially review-and-fix a change set, branch, or commit range until clean.