nano-banana (image prompt craft)

intent

write prompts that extract studio-quality images out of nano banana, google's gemini 3 pro image model (the flash variant runs faster and cheaper). use this skill when you need an ai image for social (quote graphic, carousel slide, thumbnail, infographic, product shot, ad creative) and you want legible in-image text, consistent multi-image sets, or conversational edits. nano banana excels at text rendering and style consistency; route pure typography/layout work to ideogram and surreal fine art to midjourney. reads your brand profile for colours and visual style. always discloses ai generation and carries synthid watermark (never strips it or passes ai work off as real).

inputs

brand context:

• brand-profile.md , required. colours, visual style, do/don't list.

user input:

• job type (quote graphic / carousel / thumbnail / infographic / product shot / background / ad / edit request).
• platform target and desired aspect ratio (1:1 for instagram feed, 9:16 for reels, 16:9 for youtube, etc.).
• copy/text that must appear in the image (exact wording, preferred font, placement hints).
• reference images (optional but recommended for multi-image consistency).

external connections:

• google gemini 3 pro image api (model id: gemini-3-pro-image or flash variant gemini-3-flash-image). requires GEMINI_API_KEY env var with generativelanguage.googleapis.com scope.
• woopsocial media integration (for upload and attachment flow). see tools/integrations/nano-banana.md for api setup, pricing, rate limits (100 req/min per default quota), and auth refresh cadence (token expiry every 24h).

supporting references:

• references/when-and-how-to-prompt.md , when to use nano banana vs other models and prompt anatomy.
• references/text-and-consistency.md , in-image text strategy and multi-image consistency workflow.
• references/social-recipes.md , adaptable prompt patterns per job type and aspect ratios.
• references/examples.md , weak-to-strong prompt examples, consistent carousel case study, edit example.

procedure

step 1: load brand and decode the job

input: brand-profile.md, user request (job type, platform, copy, reference images if any).

• read brand-profile for palette, typography cues, visual do/dont.
• identify the job (quote graphic, carousel, thumbnail, infographic, product, background, ad, or edit).
• note platform and aspect ratio (e.g. instagram feed = 1:1, youtube thumbnail = 1.33:1).
• note any copy that must render in the image (exact wording).

output: internal checklist of brand constraints and job scope.

step 2: write the prompt in natural language

input: brand constraints, job scope, copy, aspect ratio, any reference images.

• describe the scene in natural language: subject, composition and camera angle, action or context, lighting and atmosphere, style or medium, and constraints.
• embed exact copy in quotes (e.g. "place the text 'summer sale' in white sans-serif at the top center").
• include brand palette and style cues (e.g. "use the brand palette of deep blue #1e3a8a and cream #faf8f3").
• set aspect ratio explicitly (e.g. "aspect ratio 1:1 for instagram feed").
• do not pad with token spam ("4k, masterpiece, trending on artstation"). gemini models reason over natural language.
• keep in-image text short and verify spelling before send.

output: one natural-language prompt ready for api call.

step 3: generate at test resolution, then final

input: prompt from step 2.

• call nano banana api (gemini 3 pro image) at 1024 px for fast iteration (cheaper, ~2-5 sec).
• review for subject accuracy, composition, text legibility, brand alignment.
• if adjustments needed, rewrite prompt and regenerate at 1024 px.
• once locked, generate at 2048 or 4096 px for final quality (final images for publish).

output: hero image (1024 px iteration or 2048-4096 px final).

step 4: if generating a multi-image set, use consistency protocol

input: locked prompt from step 3, reference images, aspect ratio.

• supply reference image(s) and locked style description to nano banana.
• generate each new variant one at a time (e.g. "same scene and style, but swap the foreground subject to X").
• hold aspect ratio constant across the set.
• review each image against references and previous images for drift (colour, tone, composition, character consistency).
• if drift detected, regenerate or rewrite the variant prompt.

output: 2+ images in a consistent set.

step 5: verify, disclose, ship

input: final image(s) from step 3 or 4.

• verify all rendered text: spelling, grammar, placement, legibility. the model can misspell or misplace text.
• verify infographic data (numbers, labels, order) against source.
• confirm synthid watermark is present (do not strip it, do not pass as non-ai photo).
• check ai disclosure requirements per platform and region (meta: "made with ai", eu ai act: disclosure required).
• upload image bytes to woopsocial media via tools/integrations/nano-banana.md (api call: post to /media/upload).
• attach image to post via scheduling-and-queue skill.
• publish or schedule post.

output: image live on platform with ai disclosure visible.

decision points

when to reach for nano banana vs another model:

• if the job is legible in-image text, multi-image consistency, or conversational editing, use nano banana.
• if the job is pure typography/layout-led design (magazine cover, poster layout), route to ideogram instead.
• if the job is surreal or fine-art aesthetics, route to midjourney instead.
• if the job is photorealistic headshot or product detail, use nano banana (gemini's photoreal lane; imagen is deprecated as of aug 17 2026).

if the prompt is keyword spam:

• rewrite into a specific scene description. drop "4k, masterpiece, trending on artstation" tokens. nano banana reasons over natural language, not keyword tokens.

if the user asks for real identifiable people or copyrighted characters/logos:

• refuse and offer an original alternative (e.g. "i can't render real people or copyrighted characters. how about an illustrated person with similar vibe" or "can't use that logo; here's a similar design in brand colours").

if the user asks to post the image as a normal photo without disclosure:

• refuse. synthid watermark + ai disclosure are required per platform tos and regional law (eu ai act). never disguise ai imagery.

if in-image text is a long paragraph:

• expect errors or illegibility. shorten the copy and verify output. if the user needs design-perfect text layout, finish in a design tool (figma, canva).

if the user says "generate it in woopsocial":

• clarify: woopsocial doesn't generate images. this skill crafts the prompt, nano banana generates the image, and then you upload to woopsocial media and attach to the post.

if api rate limit hit (100 req/min default):

• queue remaining requests or wait 60 seconds. escalate to admin if production volume exceeds quota.

if gemini api key is expired or missing:

• fail gracefully with error message. check env var GEMINI_API_KEY and re-auth if needed (24h token expiry).

if network timeout or api error on image generation:

• retry once after 5 seconds. if second attempt fails, report error to user with timestamp and request id (from api response headers).

if generated image is empty or corrupted:

• regenerate with same prompt. if issue persists, rewrite prompt (may indicate conflicting or ambiguous instructions).

output contract

success looks like:

• final image file (jpg or png, 2048+ px) saved to media/nano-banana/<job-type>/<timestamp>.<ext>.
• image carries synthid watermark + c2pa metadata.
• all in-image text is spelled correctly and placed as specified.
• image aligns to brand palette and style from brand-profile.
• aspect ratio matches platform target (e.g. 1:1 for instagram feed).
• if multi-image set: all images locked to consistent style, palette, and character/composition.
• if edit: original hero image edited per user request (colour shift, seasonal variant, locale swap, etc.) with consistency maintained.
• file metadata includes prompt used, model version (gemini-3-pro-image), generation timestamp, and ai disclosure tag.
• image attached to post in woopsocial with ai disclosure visible in caption or platform disclosure field.

outcome signal

the user knows the skill worked when:

• the image renders in the post with legible, correctly spelled text.
• the image matches the brand style (colours, tone, composition).
• if a carousel or set, all images look cohesive and avoid drift.
• synthid watermark is visible on the image (or in metadata).
• ai disclosure appears per platform requirement ("made with ai" badge on meta, text disclosure in caption, etc.).
• image is live on the target platform (instagram, facebook, youtube, tiktok, etc.) or queued for scheduled publish.
• if the user requested an edit (locale swap, seasonal variant, etc.), the edit is visually different from the hero image but maintains the same style and characters.