A Developer's Guide to the Wikiprompt /dataset Endpoint

A Developer's Guide to the Wikiprompt /dataset Endpoint

Wikiprompt is a public, curated catalog of AI prompts, covering ChatGPT, Claude, Gemini, GPT Image, Midjourney, Seedance, Veo, Kling, Nano Banana, Grok and more. Instead of scraping the live site to pull that catalog programmatically, there is now a dedicated bulk export: /dataset. This post documents it the way you'd want a third-party API documented: response shapes, field semantics, pagination mechanics, caching, and error handling, with runnable examples.

Two endpoints, one job

There are exactly two URLs you need:

The manifest is metadata about the export. The data endpoint is the export itself, paginated. Both return plain JSON, no API key, no auth header, nothing to register for.

The manifest response

A GET on the dataset manifest returns something shaped like this:

{

"total_prompts": 55000,

"record_fields": [

"slug", "url", "title", "description", "content",

"category", "tags", "media", "model", "metadata",

"author", "original_source", "created_at", "updated_at"

],

"pagination": {

"endpoint": "https://www.wikiprompt.org/dataset/prompts",

"cursor_param": "after",

"limit_param": "limit",

"default_limit": 200,

"max_limit": 500

}

}

Treat total_prompts as a live, approximate count, not a fixed constant. The catalog grows daily, so pin your integration to record_fields and the pagination block rather than hardcoding the count anywhere in your code.

The data response

A GET on /dataset/prompts returns a page of records plus a cursor for the next page:

{

"count": 500,

"records": [

{

"slug": "one-brick-monumental-shadow-architectural-photo",

"url": "https://www.wikiprompt.org/one-brick-monumental-shadow-architectural-photo",

"title": "One Brick: Monumental Shadow Architectural Photo",

"description": "A single brick photographed to cast a monumental architectural shadow.",

"content": "the actual prompt text goes here, verbatim",

"category": "creative",

"tags": ["photography", "architecture", "shadow-play"],

"media": ["https://www.wikiprompt.org/media/tw/..."],

"model": "Midjourney",

"metadata": {

"media_type": "image",

"aspect_ratio": "16:9",

"style": ["minimalist", "high-contrast"],

"assessment": { "creativity": 4, "usefulness": 3 }

},

"author": "some_handle",

"original_source": "https://twitter.com/some_handle/status/...",

"created_at": "2026-03-11T00:00:00Z",

"updated_at": "2026-03-11T00:00:00Z"

}

],

"next": "https://www.wikiprompt.org/dataset/prompts?after=<cursor>&limit=500"

}

Field reference

Every record carries the same fourteen fields. The ones worth calling out:

Pagination mechanics

The endpoint uses keyset pagination, not page numbers. Two params control it:

The contract is simple: call the endpoint, read next, call next verbatim, repeat until next is null. Do not construct the after value yourself; treat it as an opaque token.

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

A full crawl in Python looks like this:

import requests

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

records = []

while url:

resp = requests.get(url, timeout=30)

resp.raise_for_status()

payload = resp.json()

records.extend(payload["records"])

url = payload["next"]

print(f"pulled {len(records)} prompts")

At 500 per page and 55,000+ records, that is roughly 110 requests for a full sync, or far fewer for an incremental one if you additionally filter client-side on updated_at.

Caching and CORS

Both endpoints are edge-cached, so repeated requests for the same page (same after value) are cheap and fast on our side, and fast for you. Access-Control-Allow-Origin: * is set on every response, so you can call this directly from browser JavaScript with fetch(), no proxy needed. There is no rate limit tied to an API key because there is no API key; be a reasonable citizen and cache the manifest locally instead of polling it on every request.

Error handling

Under load, the endpoint can return 503 with a Retry-After header (seconds to wait before trying again). Respect it:

import time, requests

def get_with_retry(url):

while True:

resp = requests.get(url, timeout=30)

if resp.status_code == 503:

wait = int(resp.headers.get("Retry-After", "5"))

time.sleep(wait)

continue

resp.raise_for_status()

return resp.json()

Any other non-200 is worth logging and stopping on rather than retrying blindly, a malformed after cursor should not happen if you're only ever following next, but defensive code should not assume that forever.

Attribution

The prompts in this dataset are aggregated from public posts by their original authors. Wikiprompt is the aggregator, not the rights holder. If you build something with this data, cite wikiprompt.org and, per record, the original_source field pointing back to the original post. There's no formal license attached beyond that: attribute the site and attribute the author.

Worth trying

Three records to sanity-check your parser against once you've got a page pulled: a data physicalization image prompt, a monumental shadow architectural photo, and a nomadic traveler character design. All three round-trip cleanly through content, media, and metadata.

If a bulk export is more than you need, right now, two lighter-weight options exist: the search API for single queries, and the MCP server if you're wiring this into an agent rather than a script. Both sit on top of the same underlying catalog as /dataset, so nothing here is a dead end if you start smaller and grow into the bulk export later.