How to Download 55,000 AI Prompts (Free, No Scraping)

How to Download 55,000 AI Prompts (Free, No Scraping)

If you have ever tried to build a prompt dataset by scraping Twitter threads, screenshotting Discord servers, or paginating through a website's HTML with a headless browser, you already know how fragile that is. Selectors change, rate limits kick in, and half the "prompts" you collect turn out to be replies about the prompt, not the prompt itself.

Wikiprompt just removed the need for any of that. The catalog, over 55,000 curated AI prompts spanning image, video and text generation across models like Midjourney, GPT Image, Veo, Kling, Nano Banana and Claude, is now available as a plain JSON dataset you can page through with curl. No API key, no login, no scraping. This post is the fastest path from "I want the data" to a local file with 55,000 records in it.

Step 1: hit the manifest

Before pulling any records, check the manifest so your script knows what it is dealing with:

curl "https://www.wikiprompt.org/dataset"

This returns a small JSON document describing the dataset: total_prompts (the current count, 55,000+), record_fields (the schema below), and the pagination scheme. Treat this endpoint as the source of truth for the total, don't hardcode a number in your pipeline.

Step 2: pull your first page

The actual records live at /dataset/prompts. Pagination is keyset-based (a cursor, not page numbers), which means it stays fast and stable even on page 200. The default page size is 200 records; you can ask for up to 500 at a time:

curl "https://www.wikiprompt.org/dataset/prompts?limit=500"

The response is a JSON object with a list of records and a next field. next is a full URL, ready to fetch as-is, that carries an after cursor pointing at the last record you just received. When there is nothing left to page through, next is null. That single field is your entire loop condition.

Step 3: follow `next` until it's null

Here is the whole pagination loop in Python. It writes every record to a local file and stops itself:

import json

import time

import urllib.request

url = "https://www.wikiprompt.org/dataset/prompts?limit=500"

out = open("wikiprompt_dataset.jsonl", "w")

count = 0

while url:

with urllib.request.urlopen(url) as resp:

data = json.loads(resp.read())

for record in data["records"]:

out.write(json.dumps(record) + "\n")

count += 1

url = data.get("next")

time.sleep(0.2)

out.close()

print(f"Downloaded {count} prompts")

The same loop in Node, if that's your stack:

let url = "https://www.wikiprompt.org/dataset/prompts?limit=500";

const fs = require("fs");

const out = fs.createWriteStream("wikiprompt_dataset.jsonl");

let count = 0;

while (url) {

const res = await fetch(url);

const data = await res.json();

for (const record of data.records) {

out.write(JSON.stringify(record) + "\n");

count++;

}

url = data.next;

}

console.log(Downloaded ${count} prompts);

Run either one and walk away. Every response is edge-cached and CORS-enabled (Access-Control-Allow-Origin: *), so this works from a server, a browser console, a notebook, or a serverless function without any special headers. On a normal connection, pulling all 55,000+ records at limit=500 takes well under a couple of minutes.

What you actually get per record

Each JSON object in the dataset has the same shape, which is what makes it useful for anything beyond one-off browsing:

That content field is the whole point: it's the ready-to-use prompt text, not a summary of it, which is what most scraped datasets get wrong.

A few records to look at first

Rather than take the schema on faith, open a few actual entries. Titan: Engineering Blueprint of a Transforming Robot is a good example of a detailed image prompt with a full metadata block. Data Physicalization shows how category and tags get assigned for a more conceptual, design-oriented prompt. And Pastel Fantasy Portrait of a Young Woman is a compact style-driven prompt worth comparing against the style field in its metadata. Pulling those three slugs out of your local JSONL file is a fast way to sanity-check your parser before you trust it on all 55,000.

If you'd rather query than download

The bulk dataset is for when you want everything, locally, once. If you only need a slice, the search API answers on-demand queries in JSON, the MCP server exposes the same catalog as tools for Claude and other agents, and llms.txt gives crawlers a map of the site. The dataset export and these live endpoints share the same underlying data, so switching between them later costs you nothing.

One rule if you reuse this

Wikiprompt aggregates prompts from public posts by their original authors; it isn't the copyright holder. If you publish something built on this dataset, credit wikiprompt.org and the original_source link on the specific records you used. That's the only condition attached to any of this.

Download it, page through it, and build something. The manifest and the next cursor are the only two things you need to remember.