Claude Nano Banana Pro skill: 10 image use cases

brand-voice

Locks the headline copy in your house voice before you hand it to the renderer.

github

Drop the rendered hero straight into the post's PR via the GitHub MCP — no manual upload step.

App icon set, four sizes, one consistent style

Generate a square app icon and three exact-pixel resizes (1024, 512, 192, 64) that stay visually coherent.

ForSolo iOS / Android devs about to ship an MVP and not paying for a designer round-trip.

The prompt

Generate a 1:1 square app icon for a focus-timer app. Soft gradient (deep indigo → magenta), centered glyph: a stylised hourglass made of two thin curves. Flat, no drop-shadow. Render at 1K. Then generate three more 1:1 versions tightened for 512, 192, and 64 px display — strip detail at the smaller sizes so the glyph stays readable.

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "1:1 app icon, focus timer, indigo-magenta gradient, hourglass glyph, flat, no shadow" \
    --filename icon-1024.png --resolution 1K
→ Saved: ./icon-1024.png (1024×1024)
# Then 3 follow-up calls with --input-image icon-1024.png and 'simplify for 512px / 192px / 64px display'
→ Saved: icon-512.png, icon-192.png, icon-64.png — all four share the same glyph silhouette

One-line tweak

For the 64 px variant always pass the 1024 as --input-image rather than re-prompting from text — the model preserves the silhouette much more reliably than re-rolling.

Pairs with

mobile-ios-design

Wraps the icon in the App Store / TestFlight asset checklist so you don't ship a half-set.

svg-precision

Vectorise the 1024 PNG into an SVG once you're happy — clean adaptive icon for Android.

Storybook illustrations with one consistent character

Eight illustrations across a children's-book story where the same character (a small fox in a yellow coat) stays recognisable page-to-page.

ForIndie authors and parents who want a printable picture book without hiring an illustrator.

The prompt

Pass 1: generate the canonical character — "a small fox in a yellow coat, watercolor style, gentle expression", 1:1, 2K, save as fox-ref.png. Then for each scene call the skill with --input-image fox-ref.png and a per-page prompt: "the fox stands at a forest crossroads at dusk", "the fox crosses a wooden bridge in the rain", etc. Keep watercolor style, 4:3 aspect, 1K.

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "the fox crosses a wooden bridge in the rain, watercolor, 4:3" \
    --input-image fox-ref.png \
    --filename page-04-bridge.png --resolution 1K
→ Saved: ./page-04-bridge.png — same fox silhouette and coat colour as fox-ref.png
  Model: gemini-3-pro-image-preview · 8 reference slots used (max 14)

One-line tweak

If face drift creeps in by page 6, stack two references on the next call — the canonical fox-ref.png plus the most recent in-style page. Pro will average the two and snap back.

Pairs with

brand-voice

Generates the per-page narration in a consistent reading-age voice.

replicate

Fallback batch lane if Pro's per-page latency stacks up over an 8-page book.

Product mockup — hero shot in real context

Drop a product (mug, t-shirt, packaged box) into a believable lifestyle scene without a photo studio.

ForE-commerce founders launching a Shopify store and skipping the $1,200 product-shoot quote.

The prompt

Use --input-image product-flat.png (the bare product on white). Generate a lifestyle shot: "the mug sits on a sun-lit oak desk with a half-open notebook and a houseplant blurred in the background, golden-hour light, photorealistic, 3:2", 2K. Keep the product geometry exactly — no logo distortion.

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "the mug sits on a sun-lit oak desk, half-open notebook, blurred houseplant, golden-hour, photoreal, 3:2" \
    --input-image product-flat.png \
    --filename mug-lifestyle-01.png --resolution 2K
→ Saved: ./mug-lifestyle-01.png (3072×2048) — logo intact, depth-of-field correct
  Tip: pass --input-image to lock product geometry; text-only prompts re-imagine the product

One-line tweak

If the model re-paints your logo, lower the prompt's scene complexity and add "do not modify the printed logo on the mug" — the constraint is respected better than you'd expect.

Pairs with

gemini-image-generation

Useful when you need the same generation to be callable from a non-Claude agent (Cursor, internal bot) sharing the billing project.

frontend-slides

Drop the lifestyle mockups straight into the pitch deck once they're rendered.

Palette swap on an existing image

Take one finished illustration and produce three brand-coloured variants (e.g. for A/B-testing landing-page hero).

ForGrowth engineers running palette tests without re-commissioning the artwork.

The prompt

--input-image hero-original.png. Generate three variants: (1) "recolour to brand palette: #0E1B3F base, #1AC4B7 accent, keep all geometry", (2) "warm sunset palette: #1B0E2A base, #FF8A3D accent", (3) "high-contrast monochrome, near-black + single neon-lime accent". 16:9, 2K each.

What slides.md looks like

$ for VARIANT in cool-teal warm-sunset mono-lime; do
    uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
      --prompt "recolour to ${VARIANT} palette, keep all geometry and composition exactly" \
      --input-image hero-original.png \
      --filename hero-${VARIANT}.png --resolution 2K
  done
→ 3 files written, geometry preserved, only palette shifted

One-line tweak

Hex codes in the prompt outperform colour-name descriptors. Pro reads "#0E1B3F" more reliably than "deep navy".

Pairs with

brand-voice

Pulls the canonical hex palette out of your brand guidelines so you're not eyeballing the codes.

github

Commits all three palette variants in one PR as A/B-test candidates.

Photoreal product shot for an e-commerce listing

A clean white-background product photo (the kind Amazon and Shopify require) from a single rough phone snap.

ForEtsy and Shopify sellers without a lightbox.

The prompt

--input-image phone-snap.jpg. "Photorealistic e-commerce product shot, pure white seamless background, soft top-down studio light, no shadow under product, 1:1 square, 2K, sharp focus on product, remove any background clutter from the source."

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "photoreal e-commerce shot, pure white seamless bg, soft top-down studio light, 1:1, sharp focus, no clutter" \
    --input-image phone-snap.jpg \
    --filename listing-front.png --resolution 2K
→ Saved: ./listing-front.png (2048×2048) — Amazon-spec compliant
  Watermark: SynthID is baked in (invisible) — that's a hard requirement to know about for stock-photo licensing

One-line tweak

Generate four shots (front, three-quarter, top-down, detail) using the same --input-image. Pro keeps the product geometry locked across all four when you reuse the reference.

Pairs with

gemini-2-5-flash-image

Switch to the Flash route for bulk variant catalogues — the original Nano Banana is faster and good enough for the 30th colour SKU.

comfyui-workflow-builder

When you need a deterministic local pipeline (Stable Diffusion + LoRA + ControlNet) instead of a hosted API call.

Architectural / interior rendering from a sketch

Turn a hand-drawn floor plan or rough perspective into a photoreal interior visualisation.

ForArchitects, interior designers, and Airbnb hosts pre-renovation.

The prompt

--input-image sketch-livingroom.jpg. "Photorealistic interior, mid-century modern living room derived from this sketch, walnut floor, off-white walls, warm afternoon light through tall windows, single large bouclé sofa, terrazzo coffee table. 16:9, 4K."

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "photoreal mid-century modern living room from this sketch, walnut floor, bouclé sofa, terrazzo coffee table, warm afternoon light, 16:9" \
    --input-image sketch-livingroom.jpg \
    --filename livingroom-render-4k.png --resolution 4K
→ Saved: ./livingroom-render-4k.png (5632×3072, ~24 MB)
  Cost: ~$0.24 per 4K render — set a Google Cloud budget alert before iterating

One-line tweak

4K is where Pro genuinely beats every other model right now, but it doubles per-image cost. Iterate at 2K, then re-render the chosen composition once at 4K.

Pairs with

excalidraw-architect

Sketch the floor-plan in Excalidraw first, export the PNG, then feed it as --input-image.

drawio-diagrams-enhanced

Same idea but for measured plans — drawio gives you the dimension annotations Pro will respect on the render.

Infographic with embedded text and data labels

A single 4:5 social-ready infographic with five labelled segments and a headline — text rendered crisply inside the image.

ForNewsletter authors, B2B marketers, and policy researchers who keep paying Canva for the same 5-segment template.

The prompt

Generate a 4:5 infographic, headline "5 Causes of MCP Server Bloat", five horizontal rows each with a 3-word label and a 12-word body, light card backgrounds on a navy gradient, accent colour #1AC4B7, render all text crisply in Inter, 2K.

What slides.md looks like

$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "4:5 infographic, headline '5 Causes of MCP Server Bloat', 5 rows, navy gradient, accent #1AC4B7, all text crisp in Inter" \
    --filename infographic-mcp-bloat.png --resolution 2K
→ Saved: ./infographic-mcp-bloat.png (1638×2048) — 5 labels + body copy all legibly rendered
  Note: text rendering is the killer feature here — no other API model nails 60+ characters of in-image type

One-line tweak

Quote your label and body copy verbatim inside the prompt with quotes — Pro respects literal strings far more reliably than paraphrased descriptions.

Pairs with

drawio-diagrams-enhanced

When the data is structured (org chart, system diagram), drawio + Pro gives you a labelled diagram + an editorial render of the same thing.

frontend-slides

Embed the rendered infographic into a deck slide directly.

Brand-coloured marketing banner batch

Five 1200×628 (1.91:1) social banners — one per platform — using the same brand palette and a rotating tagline.

ForSolo marketers running a launch week across LinkedIn, X, Facebook, Instagram, Threads.

The prompt

Generate five banners at 16:9 (then crop to 1.91:1 with Pillow). Brand palette: #0E1B3F base, #1AC4B7 accent, white text. Each banner shares the product wordmark on the left and rotates the tagline: "Ship faster" / "Ship safer" / "Ship together" / "Ship smarter" / "Ship calmly". 2K each.

What slides.md looks like

$ for TAG in "Ship faster" "Ship safer" "Ship together" "Ship smarter" "Ship calmly"; do
    uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
      --prompt "16:9 marketing banner, palette #0E1B3F + #1AC4B7, wordmark left, tagline right: '${TAG}'" \
      --filename banner-${TAG// /-}.png --resolution 2K
  done
# Then crop to 1.91:1 (1200×628) — the only common social ratio Pro doesn't natively output
$ python -c "from PIL import Image; [Image.open(f).crop((0, 224, 3072, 1840)).save(f) for f in __import__('glob').glob('banner-*.png')]"

One-line tweak

1.91:1 (LinkedIn / Open Graph) isn't in Pro's native aspect-ratio list — generate at 16:9 and centre-crop to 1.91:1 with Pillow. Cropping is lossless from a 2K source.

Pairs with

brand-voice

Generate the rotating tagline in your brand voice before piping it in — you'll catch tonal misfires before the render burn.

replicate

If you need the same five banners in 25 language variants for international launch, Replicate's batch route is cheaper than Pro per call.

Logo concept sheet with iterative refinement

Six logo concepts on one sheet, then iterate the chosen concept through three refinement passes (simpler, stronger contrast, monogram-only).

ForFounders pre-brand-identity work, agencies pitching concept sheets, and side-project naming sessions.

The prompt

Pass 1: "Generate a 6-cell concept sheet for a logo for 'Loomwave', a real-time data-streaming startup. Six distinct directions: geometric, wordmark, monogram, abstract symbol, line-illustration, badge. 2:3, 2K." Pass 2: pick concept #4 by passing it as --input-image and prompting "isolate concept 4 only, simplify, increase stroke weight, 1:1". Pass 3: "reduce to monogram letterform only, single accent colour".

What slides.md looks like

# Pass 1 — concept sheet
$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "6-cell logo concept sheet for 'Loomwave', directions: geometric, wordmark, monogram, abstract, line-illustration, badge" \
    --filename loomwave-sheet.png --resolution 2K
→ Saved: ./loomwave-sheet.png

# Pass 3 — final monogram, refined twice
$ uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
    --prompt "reduce to monogram letterform only, single #1AC4B7 accent, 1:1" \
    --input-image loomwave-iter-2.png \
    --filename loomwave-final.png --resolution 2K

One-line tweak

If the iteration drifts an attribute you wanted to keep (e.g. you asked for thicker strokes and the colour also shifted), split the prompt into two single-edit passes — Pro respects single-axis edits much more cleanly than compound ones.

Pairs with

svg-precision

Vectorise the final monogram once you're happy — you need a clean SVG for the favicon and the brand kit.

nano-banana-pro-prompts-recommend-skill

If you blanked on the six 'directions' for the concept sheet, this companion skill recommends prompts from a 10k+ catalogue.

Community signal

Three voices from people putting Gemini 3 Pro Image through real workloads. The first is the clearest comparative endorsement; the second is the daily-use story; the third is the launch-day pricing pushback that every team feels on day two.

“Nano Banana Pro aka gemini-3-pro-image-preview is the best available image generation model. … this is an astonishingly capable image generation model.”

Simon Willison · Blog

Launch-day review on simonwillison.net. Willison tested 4K output (a 24.1 MB 5632×3072 image) and infographic text rendering before declaring it the new default.

“Nano Banana Pro drops today! I've had about a week with it now and it's really impressive. Combine up to 6 images into a single image. … The best text rendering in an image generator we've seen so far.”

Matt Wolfe (@mreflow) · X / Twitter

Pre-launch impressions thread on X from a creator with early access. The text-rendering line became the most-quoted endorsement on launch day.

“One annoying aspect of the thinking step is that it makes generation time inconsistent: I've had 2K generations take anywhere from 20 seconds to one minute, sometimes even longer during peak hours.”

Max Woolf (minimaxir) · Blog

December 2025 deep-dive review. Woolf is otherwise very positive on quality but lands hard on the latency/cost trade — the part every team feels on day two.

The contrarian take

Not everyone is keeping Pro as their default. The most precise criticism on the launch threads is from Max Woolf:

“The increased cost and generation time is a severe constraint on many fun use cases outside of one-off generations.”

Max Woolf · Blog

From Max Woolf's December 2025 deep-dive review.