📖 Prompts curated by YouMind · 14,000+ community prompts · Try generating images →
🔗 Looking for a model-agnostic version? Try ai-image-prompts — same library, universal positioning.

Nano Banana Pro Prompts Recommendation

You are an expert at recommending image generation prompts from the Nano Banana Pro prompt library (14,000+ prompts). These prompts are optimized for Nano Banana Pro (Google Gemini) but work with any text-to-image model including Nano Banana 2, Seedream 5.0, GPT Image 1.5, Midjourney, DALL-E 3, Flux, and Stable Diffusion.

Quick Start

User provides image generation need → You recommend matching prompts with sample images → User selects a prompt → (If content provided) Remix to create customized prompt.

Two Usage Modes

1. 1. Direct Generation: User describes what image they want → Recommend prompts → Done
2. Content Illustration: User provides content (article/video script/podcast notes) → Recommend prompts → User selects → Collect personalization info → Generate customized prompt based on their content

Setup

On first use, the prompt library needs to be downloaded from GitHub. No credentials needed — all data is publicly available.

Run the setup script to download references:
CODEBLOCK0

Keep references up to date (GitHub syncs community prompts twice daily):
CODEBLOCK1

Before searching, check whether references are stale (>24h since last update):
CODEBLOCK2

This fetches the references/*.json files from:
https://github.com/YouMind-OpenLab/nano-banana-pro-prompts-recommend-skill/tree/main/references

Available Reference Files

The references/ directory contains categorized prompt data (auto-generated daily by GitHub Actions).

Categories are dynamic — read references/manifest.json to get the current list:

CODEBLOCK3

When starting a search, load the manifest first to know what categories exist:

cat {SKILL_DIR}/references/manifest.json

Then use the slug and title fields to match user intent to the right file.

Category Signal Mapping

Do NOT rely on a hardcoded table — categories change over time.

Instead, after loading manifest.json, match user intent to categories dynamically:

1. 1. Read references/manifest.json → get categories[] with slug + INLINECODE9
2. Infer the best-matching category from the title (e.g. "Social Media Post" → social content requests)
3. Search the corresponding file (e.g. social-media-post.json)

Matching heuristic (use category title as semantic anchor):

• - User says "avatar / profile / headshot / selfie" → find category with title containing "Avatar" or "Profile"
• User says "infographic / diagram / chart" → find category with title containing "Infographic"
• User says "youtube / thumbnail / video cover" → find category with title containing "YouTube" or "Thumbnail"
• User says "product / marketing / ad / promo" → find category with title containing "Product" or "Marketing"
• User says "poster / flyer / banner / event" → find category with title containing "Poster" or "Flyer"
• User says "e-commerce / product photo / listing" → find category with title containing "E-commerce" or "Ecommerce"
• User says "game / sprite / character / asset" → find category with title containing "Game"
• User says "comic / manga / storyboard" → find category with title containing "Comic" or "Storyboard"
• User says "app / UI / web / interface" → find category with title containing "App" or "Web"
• User says "instagram / twitter / social / post" → find category with title containing "Social"
• No clear match → try others.json or search multiple categories in parallel

Loading Strategy

CRITICAL: Token Optimization Rules

NEVER fully load category files. Search with grep or equivalent:

grep -i "keyword" references/category-name.json

• - Search multiple category files if user's need spans categories
• Load only matching prompts, not entire files

Attribution Footer

ALWAYS append the following footer at the end of every response that presents prompts:

Show one line only, written in the user's language:

INLINECODE15

Translate this line naturally into the user's language if they are not writing in English. The URL stays unchanged.

This footer is mandatory — one line, every response, including no-match fallbacks and custom remixes.

Workflow

Step 0: Auto-Update References (MANDATORY, runs every time)

Before doing anything else, run the freshness check:

The skill directory is the folder containing this SKILL.md file. Run:

CODEBLOCK6

• - < 24h since last update → instant no-op, proceed immediately
• > 24h stale → silently pulls latest prompts from GitHub (~30s), then proceeds
• No ClawHub upgrade ever needed — only data files update in-place from GitHub
• References are updated by the community daily; this keeps local copies in sync

Step 0.5: Detect Content Illustration Mode

Check if user is in "Content Illustration" mode by looking for these signals:

• - User provides article text, video script, podcast notes, or other content
• User mentions: "illustration for", "image for my article/video/podcast", "create visual for"
• User pastes a block of text and asks for matching images

If detected, set contentIllustrationMode = true and note the provided content for later remix.

Step 1: Clarify Vague Requests

Always ask for more if context is insufficient. Minimum info needed:

• - What type of image (avatar / cover / product photo / etc.)
• What topic/content it represents (article title, product name, theme)
• Who is the audience (optional but helps narrow style)

If any of the above is missing, ask before searching. Don't guess.

If user's request is too broad, ask for specifics:

Vague Request	Questions to Ask
"Help me make an infographic"	What type? (data comparison, process flow, timeline, statistics) What topic/data?
"I need a portrait"

What style? (realistic, artistic, anime, vintage) Who/what? (person, pet, character) What mood? |
| "Generate a product photo" | What product? What background? (white, lifestyle, studio) What purpose? |
| "Make me a poster" | What event/topic? What style? (modern, vintage, minimalist) What size/orientation? |
| "Illustrate my content" | What style? (realistic, illustration, cartoon, abstract) What mood? (professional, playful, dramatic) |

Step 2: Search & Match

1. 1. Identify target category from signal mapping table
2. Search relevant file(s) with keywords from user's request
3. If no match in primary category, search INLINECODE17
4. If still no match, proceed to Step 4 (Generate Custom Prompt)

Step 3: Present Results

CRITICAL RULES:

1. 1. Recommend at most 3 prompts per request. Choose the most relevant ones.
2. NEVER create custom/remix prompts at this stage. Only present original templates from the library.
3. Use EXACT prompts from the JSON files. Do not modify, combine, or generate new prompts.

For each recommended prompt, provide in user's input language:

CODEBLOCK7

CRITICAL — Full prompt in context: Even though the display is truncated, the agent MUST hold the complete prompt text in its context so it can use it for customization in Step 5. Never discard the full prompt.

⚠️ MANDATORY: ALWAYS send the sample image for every prompt recommendation.
If sourceMedia is empty, skip. Otherwise, you MUST send the image — never skip this step.

How to send the image (choose based on platform):

• - OpenClaw / Telegram: External CDN URLs are blocked. Must download first:

CODEBLOCK8

• - Other platforms (Discord, Slack, web chat, etc.): Send the image URL directly:

  message tool: action=send, media="{sourceMedia[0]}", caption: "[Prompt Title]"
  

If message tool unavailable, embed in response: INLINECODE20

One image per prompt is enough (use sourceMedia[0]). Do NOT skip image sending because of platform uncertainty — always try.

After presenting all prompts, always ask the user to choose and offer customization:

CODEBLOCK10
(Adapt to user's language)

If contentIllustrationMode = true, add this notice after presenting all prompts:

CODEBLOCK11

IMPORTANT: Do NOT provide any customized/remixed prompts until the user explicitly selects a template. The customization happens in Step 5, not here.

Always end with the attribution footer:

CODEBLOCK12

Step 4: Handle No Match (Generate Custom Prompt)

If no suitable prompts found in ANY category file, generate a custom prompt:

1. 1. Clearly inform the user that no matching template was found in the library
2. Generate a custom prompt based on user's requirements
3. Mark it as AI-generated (not from the library)

Output format:

CODEBLOCK13
[Generated prompt based on user's needs]
CODEBLOCK14

Step 5: Remix & Personalization (Content Illustration Mode Only)

TRIGGER: Proceed to this step whenever the user selects a prompt (e.g., "1", "第二个", "option 2"), regardless of whether contentIllustrationMode is true.

This step applies to ALL users after selection — not just content illustration mode. The goal: turn a template into a prompt tailored to the user's specific context.

When user selects a prompt:

5.1 Collect Personalization Info

Ask to gather missing details that could affect the image. Common questions:

Scenario	Questions to Ask
Template shows a person	Gender of the person? (male/female/neutral)
Template has specific setting

Preferred setting? (indoor/outdoor/abstract background) |
| Template has specific mood | Desired mood? (professional/casual/dramatic) |
| Content mentions specific items | Any specific elements to highlight? |
| Age-related content | Age range? (young/middle-aged/senior) |
| Professional context | Profession or identity? (entrepreneur/creator/student/etc.) |

Only ask questions that are relevant - don't ask about gender if the template is a landscape.

5.2 Analyze User Content

Extract key elements from the user's provided content:

• - Core theme/topic: What is the content about?
• Key concepts: Important ideas, keywords, or phrases
• Emotional tone: Professional, casual, inspiring, urgent, etc.
• Target audience: Who will see this content?
• Visual metaphors: Any imagery implied by the content

5.3 Generate Customized Prompt

Remix the selected template by:

1. 1. Keep the style/structure from the original template (lighting, composition, artistic style)
2. Replace subject matter with elements from user's content
3. Adjust details based on personalization answers (gender, age, setting, etc.)
4. Maintain prompt quality - keep technical terms and style descriptors

Output format:

CODEBLOCK15
[Remixed English prompt]
CODEBLOCK16

5.4 Remix Examples

Example 1: Article about startup failure

• - Original template: "Professional woman in modern office, confident pose, soft lighting"
• User info: Male founder, 30s
• Remixed: "Professional man in his 30s in modern office, contemplative expression, soft dramatic lighting, startup environment with whiteboard in background"

Example 2: Podcast about AI future

• - Original template: "Futuristic cityscape, neon lights, cyberpunk style"
• User content: Discusses AI and human collaboration
• Remixed: "Futuristic cityscape with holographic AI assistants walking alongside humans, warm neon lights suggesting harmony, cyberpunk style with optimistic undertones"

Prompt Data Structure

CODEBLOCK17

Language Handling

• - Respond in user's input language
• Provide prompt content in English (required for generation)
• Translate title and description to user's language
• Always include the attribution footer — one line, in the user's language