Agent-Skills.md

Agent Skills: Gemini Image Generation & Editing (Nano Banana family)

Generate and edit images with Google's Gemini image models (Nano Banana family) via a bundled script. Text-to-image, image editing, and multi-reference composition at 1K/2K/4K with aspect-ratio control across lite/flash/pro model tiers. Do NOT read image files and do NOT write new image-generation code; always call the bundled script, passing image paths via --input-image.

UncategorizedID: garg-aayush/tutorials/gemini-image

Repository

garg-aayush/tutorials
garg-aayushLicense: Apache-2.0
93

Install this agent skill to your local

pnpm dlx add-skill https://github.com/garg-aayush/handbook/tree/HEAD/.claude/skills/gemini-image

Skill Files

Browse the full folder contents for gemini-image.

Download Skill

Loading file tree…

.claude/skills/gemini-image/SKILL.md

Skill Metadata

Name
gemini-image
Description
Generate and edit images with Google's Gemini image models (Nano Banana family) via a bundled script. Text-to-image, image editing, and multi-reference composition at 1K/2K/4K with aspect-ratio control across lite/flash/pro model tiers. Do NOT read image files and do NOT write new image-generation code; always call the bundled script, passing image paths via --input-image.

Gemini Image Generation & Editing (Nano Banana family)

Generate new images or edit existing ones with Google's Gemini image models. Runs non-interactively in a fork: generate, save, and report the path(s) back.

Operating loop

  1. Classify the request: generate, edit (one input image), or multi-reference composition (several input images).
  2. For crafted output (posters, text-heavy images, product renders, diagrams, photorealism) read the relevant section of references/craft.md first; load the smallest useful slice.
  3. Pick model, resolution, and aspect from the policy tables below.
  4. Pre-flight: input image paths exist; run from the user's working directory so output lands where they work.
  5. Run the script, then report the Saved: path(s) plus one concrete refinement suggestion.
  6. If the request is ambiguous AND expensive (4K, pro model, or a big batch), do not spend the credits: return 1-3 proposed directions with the flags you would use, and let the caller pick.

Usage

# Generate
uv run ${CLAUDE_SKILL_DIR}/scripts/generate_image.py --prompt "description" --name "short-slug"

# Edit an existing image
uv run ${CLAUDE_SKILL_DIR}/scripts/generate_image.py --prompt "editing instructions" --name "slug" --input-image path/to/input.png

# Multi-reference composition (repeat --input-image; index them in the prompt)
uv run ${CLAUDE_SKILL_DIR}/scripts/generate_image.py --prompt "Put the product from Image 1 into the scene from Image 2" -i product.png -i scene.jpg --name "composite"

# Variants
uv run ${CLAUDE_SKILL_DIR}/scripts/generate_image.py --prompt "..." --name "logo-ideas" -n 4

Model policy

| Alias | Model ID | Use for | |---|---|---| | flash (default) | gemini-3.1-flash-image | Everything unless stated otherwise; 4K-capable generalist, up to 14 reference images | | pro | gemini-3-pro-image-preview | Final assets, dense/accurate text rendering, brand precision, complex multi-step edits, or when the user asks for top quality | | lite | gemini-3.1-flash-lite-image | Bulk drafts and thumbnails; 1K only (script auto-degrades) | | legacy | gemini-2.5-flash-image | Only if the user explicitly asks |

--model also accepts raw model IDs, so new models work without a script change. If the API returns model-not-found, the preview alias likely graduated: check https://ai.google.dev/gemini-api/docs/models and pass the current ID directly.

Resolution and aspect policy

Resolution (--resolution, uppercase K): default 1K; "2K/2048/medium" gets 2K; "high-res/4K/ultra" gets 4K. When editing, the script auto-matches the largest input image unless the user set a resolution explicitly. Draft at 1K, finalize at 2K/4K.

Aspect (--aspect): omit for the model default. Shortcuts: square (1:1, social/avatars), portrait (3:4, posters/phone), landscape (4:3), wide (16:9, hero images/slides), tall (9:16, stories/mobile), banner (21:9). Raw ratios like 3:2 pass through.

Filename handling

Omit --filename; the script stamps the current time itself and writes <yyyy-mm-dd-hh-mm-ss>-<slug>.<ext>.

  • --name: short descriptive slug, 1-5 lowercase hyphen-separated words from the user's prompt (e.g. "A serene Japanese garden" gets --name japanese-garden). Defaults to image.
  • --filename: only when the user explicitly names the output file (overrides --name).
  • --format png|jpeg|webp (+ --compression 0-100 for jpeg/webp) when the user wants a specific format or smaller files.
  • With -n N variants, files get _0, _1, ... suffixes.

Editing

Pass editing instructions in --prompt with the input via --input-image. State the transformation first, then pin the invariants: "change only X; keep everything else exactly the same". Repeat the invariants on every iteration; make one change per iteration. See references/craft.md for the full pattern.

API key

Resolution order: --api-key argument, then GEMINI_API_KEY env var. Missing key exits with code 2 and instructions; relay them and suggest exporting GEMINI_API_KEY in ~/.zshrc. Never print the key value.

Output and exit codes

  • Saved paths are printed as Saved: <absolute path>; report these verbatim. Do not read the image back.
  • Exit 0 = success; 1 = API error or refusal (stderr has the verbatim error; surface it, especially moderation refusals on real-person edits); 2 = bad arguments or missing key (fix and retry, no credits were spent).

References

Load only the section needed; never all files by default.

  • references/craft.md: battle-tested prompt patterns, posters/text, product renders, photorealism, diagrams, UI mockups, edit invariants, multi-reference composition.
  • references/gemini-guide.md: Google's official fill-in-the-blank prompt templates (7 categories) and editing patterns.
  • references/gallery.md: 20 curated, attributed community prompts across 8 categories, each with suggested flags and the reusable pattern it demonstrates.