Grepping 10,000+ Nano Banana Pro prompts from inside a Claude Code skill

One sentence in, three prompts out

The problem is narrow and real: writing a good prompt for an image model takes longer than generating the image, and phrasing that works for one subject rarely transfers to another. This project answers that by handing an assistant a searchable corpus instead of a blank prompt box. You describe an outcome in ordinary language and the assistant returns up to 3 prompts, each with a translated title and description, the exact English text to copy, sample images to preview the style, and a note on whether reference images are needed.

There is no API key, no server and no image generation inside the repository. The skill produces a prompt string, and the image comes from whatever model you paste it into. Nano Banana Pro is Google's Gemini image model, and the documentation quotes its model ID as gemini-3-pro-image-preview. The intended user is someone already working inside Claude Code, OpenClaw, Cursor or a similar tool, who would rather describe an outcome than browse a gallery.

grep over JSON keeps the context window small

Mechanically the assistant reads SKILL.md, works out which category a request belongs to from keyword signals, then runs a grep style search over the matching JSON file under references/ and returns the top 3 hits with their images and descriptions. No full category file is ever loaded, and that is the whole point: pulling 10,000+ prompts into the context window would cost more than the image generation it is meant to help with. Rebuilding those reference files is a separate maintenance task, exposed through npx tsx scripts/generate-references.ts.

User describes need
      ↓
Skill identifies category from keyword signals
      ↓
Searches matching JSON file (token-efficient grep, never loads full file)
      ↓
Returns top 3 prompts with images + translated descriptions
      ↓
[Optional] User picks one → Skill remixes it to match their content

Two modes sit on that flow. Direct search is the one above. Remix mode takes pasted content instead of a description, an article, a video script or podcast notes, then proposes style templates, asks a few questions about gender, mood and setting, and writes a customized prompt for your draft. That clarifying round trip is the cost of the mode: you answer two or three questions and get a prompt shaped to the content you already had.

clawhub install versus the npx skills installer

OpenClaw is the documented first path, and it carries the shortest command in the project:

clawhub install nano-banana-pro-prompts-recommend

Claude Code and everything else share one installer. npx skills i is described as a universal installer that auto detects your AI assistant, which is how a single command covers Cursor, Codex, Gemini CLI and Windsurf:

npx skills i YouMind-OpenLab/nano-banana-pro-prompts-recommend-skill

A third route exists for manual setups, and OpenClaw chat search is offered as an alternative to typing the command yourself:

npx openskills install YouMind-OpenLab/nano-banana-pro-prompts-recommend-skill

Nothing gets compiled here. The primary language is TypeScript, but the TypeScript is maintenance tooling rather than a runtime: scripts/generate-references.ts rebuilds the references, and tsconfig.json with tsc --noEmit only matters if you plan to change that generator. As a first real use, ask in plain chat and compare what comes back against what you would have written by hand:

"Find me a cyberpunk-style avatar prompt"
"I need prompts for travel blog article covers"
"Looking for a product photo on white background"

The library lands on disk through a postinstall hook

Installation is not self contained, and that is the part to read before typing any command. package.json runs node scripts/setup.js on postinstall, and the package description calls that an automatic download of the library. The same script is reachable later without reinstalling, and a forced refresh is a third entry:

  "scripts": {
    "generate": "npx tsx scripts/generate-references.ts",
    "typecheck": "tsc --noEmit",
    "postinstall": "node scripts/setup.js",
    "setup": "node scripts/setup.js",
    "sync": "node scripts/setup.js --force"
  },

So installing this skill executes a Node script that downloads data into your checkout, and no offline or vendored install is documented for anyone who needs one. Runtime dependencies are modest, dotenv and qs-esm, while the development side adds tsx, typescript and @types/node under packageManager [email protected]. On permissions, the README badge points at a LICENSE file that is not among the top level repository entries, while package.json declares "license": "MIT".

The category table and its own headline disagree

The category table does not add up to its own headline. One row, Social Media Post, is listed as 10,000+, and the ten rows beneath it claim 3,600, 1,000, 900, 470, 450, 370, 370, 280, 170 and 160 between them. Whatever the true size of the corpus, those two statements cannot both describe one partition of it, because the ten smaller rows on their own already outrun the headline figure.

That imbalance matters more for a search skill than for a gallery page. The category is chosen from keyword signals before any file is searched, so a request that lands in a thin bucket can only return what that bucket holds. App / Web Design is listed at 160+ and YouTube Thumbnail at 170+, set against a 10,000+ headline row for social posts. If your subject sits in one of the smaller categories, the top 3 results come from a narrow pool, and nothing in the documentation promises a fallback search across categories when a bucket runs thin.

The README stops mid sentence on other models

The library is packaged for several models, but the support claim and the documentation are not equally strong. The package.json description says the skill also works with Nano Banana 2, Seedream 5.0, GPT Image 1.5, Midjourney, DALL-E, Flux and Stable Diffusion. The README's own answer to whether it works with other image generation models ends at the words "The prompts are optimized" and stops there, with no model specific guidance and no migration notes anywhere else in the repository.

Treat cross model support as asserted rather than demonstrated. Prompts written for one image model transfer imperfectly to another, and the skill has no per model rewriting step; only remix mode does rewriting, and only toward the content you paste. The single model identifier in the documentation, gemini-3-pro-image-preview, is itself a preview identifier, so the target can change underneath the skill without the package version moving, which sits at 1.5.5.

Twice daily curation is the only quality control

Nobody curates this by hand, and that is the difference between a prompt library you can rely on and one you cannot. The prompts are collected from viral posts by top AI artists on Twitter/X and pushed to the repository twice daily, at 00:00 and 12:00 UTC, through automated GitHub Actions. Contributions travel the same road: you post to the YouMind community and the next sync picks them up. No review step between publication and ingestion is described anywhere.

The last push was on 2026-09-20 and the repository has no GitHub releases, so the 1.5.5 in package.json has no matching release tag to pin against. Because the sync overwrites rather than appends, a prompt that worked for you can change or vanish between two runs, and the only documented remedy is running node scripts/setup.js --force to refresh a local copy. For a team that depends on a stable prompt set, copying the prompts you liked into your own file is the safer habit.

Where the gallery page beats the skill

The honest alternative is to skip the skill entirely and browse the gallery that the project links from every README section, a static page of the same prompts at youmind.com/nano-banana-pro-prompts. You lose nothing that matters when you already know the style you want. In exchange you avoid three costs: an install step, a postinstall script that downloads data, and assistant context spent on searching. What you give up is the part the skill exists for: describing an outcome in a single sentence and receiving three ranked candidates you did not know to look for.

A second, cheaper option covers repeated work. Keep a small personal file of the prompts that already worked for your subject, because a fixed library cannot invent coverage it never received, and a file you control cannot be overwritten by the next sync. Choose the skill when the use case is open ended and you are already inside an assistant. Choose the gallery, or your own file, when the subject is narrow, the brand style is fixed, or the same wording has to come out right every week.

Editorial conclusion

Adopt this skill if you make images from inside an assistant and describe outcomes rather than styles; you get three ranked prompts you would not have written, plus a remix step that asks two or three questions first. Leave it alone if you need guaranteed coverage of a narrow subject, since the library is a twice daily copy of community posts and categories as small as 160 entries exist. Before installing, read what node scripts/setup.js fetches during postinstall and keep your own copy of any prompt you plan to reuse.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. YouMind-OpenLab/nano-banana-pro-prompts-recommend-skill on GitHub