The Largest Open Dataset of AI Image and Video Prompts

The Largest Open Dataset of AI Image and Video Prompts

Most prompt collections you find online cover one model, one media type, and give you a page to scroll through by hand. A gallery of Midjourney prompts here, a gist of ChatGPT system prompts there, a paid course bundling "50 Veo prompts" behind a checkout page. None of them let you actually work with the data.

Wikiprompt is different in a way that is easy to state and worth spelling out: it is a single, open catalog of over 55,000 prompts spanning image generation, video generation, and text, covering the models people are actually using right now, GPT Image, Midjourney, Seedance, Veo, Kling, Nano Banana, ChatGPT, Claude, Grok, and more, and as of this month all of it is downloadable as a plain JSON dataset. No login, no API key, no per-model silo. One feed.

Breadth is the point

Most "prompt library" sites pick a lane. A Midjourney-only gallery cannot tell you how a Kling prompt for the same concept differs, and a text-prompt repo for ChatGPT has nothing to say about aspect ratio or style tags. Wikiprompt's catalog is organized so that image, video, and text prompts live in the same schema, browsable through the same creative prompts category and searchable through the same search API.

Concretely, that means you can pull:

That breadth is the whole thesis. If you are evaluating prompt resources to build a tool, train a classifier, or just understand what good prompting looks like across modalities, a single-model list forces you to go collect the rest yourself. Wikiprompt's dataset already has it in one place.

Every prompt ships with its result, not just its text

The other gap in scattered prompt collections is that the prompt is usually all you get. You see the text, maybe a screenshot pasted into a tweet, and you are on your own to judge whether it actually works.

Every record in the Wikiprompt dataset carries the prompt text alongside the actual media it produced, plus structured metadata: the model used, aspect ratio, resolution, style tags, and an editorial quality assessment (creativity, usefulness, technical quality, and for image/video, prompt adherence and "wow factor"). That is the difference between a list of prompt strings and a dataset you can actually use to study what works. You are not guessing whether a prompt is good, you can see the output it generated and how it scored.

How it compares to what else is out there

Scraping X or Reddit for prompts yourself means dealing with rate limits, inconsistent formatting, and posts that vanish. Paywalled "prompt pack" products give you a curated slice but lock you out of the rest and rarely include the actual generated media. Single-model communities are useful but structurally can't answer cross-model questions, like whether a particular framing style translates from Midjourney to Kling.

The Wikiprompt dataset sidesteps all three problems: it is free, it is not scoped to one model, and every record already comes with its original source linked back to the creator (the original_source field), so attribution is built in rather than bolted on.

Pulling the data

The manifest lives at the dataset, which returns the total prompt count, the record schema, and the pagination scheme. The actual catalog is at /dataset/prompts, paginated with a keyset cursor, up to 500 records per page:

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

Each response includes a next URL. Follow it until next comes back null and you have the full catalog. A minimal pagination loop looks like this:

import requests

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

records = []

while url:

resp = requests.get(url).json()

records.extend(resp["prompts"])

url = resp.get("next")

print(len(records), "prompts collected")

No API key, CORS enabled, and the whole thing sits behind Vercel's edge cache, so pulling the full set is fast and does not put load on the live site the way scraping would.

Structured fields, not free text

Each record includes slug, url, title, description, content (the actual prompt), category, tags, media, model, metadata, author, original_source, created_at, and updated_at. Categories span creative, marketing, personal, productivity, coding, education, business, research, and other, so filtering by use case is a field lookup, not a text classifier you have to train yourself.

Only active, cleaned prompts make it into the export, which means you are not downloading spam, duplicates, or half-finished drafts alongside the real thing.

If you need it live instead of in bulk

The dataset is for bulk analysis. If instead you want to query on demand, the search API returns JSON for a live query, the MCP server lets an AI agent search and pull prompts directly inside a conversation, and llms.txt gives any LLM a map of the site's structure. All three point back to the same underlying catalog as the dataset.

Attribution

Every prompt in Wikiprompt was aggregated from a public post by its original author. When you reuse a prompt or a dataset export in something you publish, credit wikiprompt.org and the record's original_source link back to the creator who wrote it. That is not a formal license, it is the baseline courtesy the whole catalog depends on.

If you have been stitching together prompt examples from bookmarked tweets and paywalled packs, this is the shortcut: one download, every modality, real outputs attached, updated as the catalog grows.