Why We Made Our Entire Prompt Catalog Open

Why We Made Our Entire Prompt Catalog Open

Every week, someone writes a scraper to pull prompts off wikiprompt.org. We can see it in the logs: a burst of requests against /category/creative, then /category/coding, working through every page of the site like a person clicking "next" a thousand times, except it is a script and it never gets bored. We used to treat this as a nuisance to rate-limit away. Then we asked ourselves a better question: why are we making people do this at all?

The answer, once we admitted it, was not flattering. We were sitting on 55,000+ curated prompts, useful raw material for anyone studying how people actually talk to AI models, and the only sanctioned way to get more than one at a time was to click through the website like it was 2004. That is not how open knowledge is supposed to work, and we know exactly what open knowledge is supposed to look like, because the best example of it has existed since 2001.

The Wikipedia model, taken seriously

Wikipedia does not just let you read articles one at a time in a browser. It publishes full database dumps of the entire encyclopedia, on a schedule, for free, no account required. Anyone can download the whole thing and do something with it: build a search engine, train a model, run an offline mirror in a place with no internet, or just study how an article changed over ten years of edits. Wikipedia did not do this reluctantly. It did it because the dump *is* the mission. An encyclopedia that only exists behind a scraping-resistant frontend is not actually open, no matter what license is printed at the bottom of the page.

The side effect nobody plans for but everyone benefits from: because the dump exists, nobody has to scrape Wikipedia to get its content at scale. The scraping traffic that would otherwise hammer the live site gets absorbed by a static file instead, and Wikipedia gets cited everywhere as a matter of course, because citing it is easier than not citing it. Openness did not weaken Wikipedia's position as the canonical source. It cemented it.

We think prompts deserve the same treatment, for the same reason. A good prompt is not really "content" in the way a blog post is content. It is a technique, a small piece of applied craft, usually built by trial and error and shared because someone wanted others to benefit from what they figured out. That is a lot closer to a Wikipedia article than to a piece of proprietary IP. A prehistoric creature sovereign portrait or an editorial portrait staged on an architectural staircase is somebody's discovery about how to talk to an image model, written down so the next person does not have to rediscover it from scratch. Locking that behind a walled garden, a login wall, or a "contact us for API access" form is the opposite of what the content is for.

Collective knowledge does not scale one request at a time

There is a specific failure mode we wanted to avoid: prompt marketplaces where the actual prompt text is hidden until you pay, or communities where the good stuff lives in a Discord server nobody outside the group can search. Those models optimize for capturing value at the point of a single transaction. We think they lose the bigger prize, which is becoming the place people go by default when they want to know how a good prompt is built, the way Wikipedia became the place you go by default for a fact. You cannot become that reference point if your content cannot be studied at scale, indexed properly, or built on top of.

So the dataset works the way a dump should work: no API key, no rate-limit tier to negotiate, CORS wide open, and cached hard enough at the edge that heavy use does not cost us sleep. Hit https://www.wikiprompt.org/dataset and you get a manifest with total_prompts and the pagination scheme. Hit `/dataset/prompts` and you get the catalog itself as JSON, paginated with a keyset cursor, up to 500 records per page:

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

Follow each response's next URL until it comes back null and you have walked the whole catalog. Every record carries slug, url, title, description, content (the prompt text itself), category, tags, media, model, structured metadata, author, original_source, and timestamps, across categories from creative and coding to marketing, research, and education. If you would rather query than bulk-download, there is also a search API and an MCP server for agents that want to reach into the catalog live instead of pulling a static copy.

Openness is not the same as anonymity

Here is the part that matters as much as the technical access: none of this is ours to strip of credit. Every prompt in the catalog traces back to a real person who posted it publicly, whether that was a defeated white snake spirit standing in the rain shared on Twitter or a technique dropped in a subreddit thread. original_source is not a decorative field. It is the other half of the deal. We are not claiming a formal license we do not hold, and we are not pretending the prompts are ours to relicense however we like. We are asking, plainly, that if you build something with this dataset, you attribute wikiprompt.org as the aggregator and the original author as the source. That is the same bargain Wikipedia strikes with everyone who reuses an article: take the knowledge, keep the credit attached.

Wikipedia's dumps did not make Wikipedia less essential. They made it infrastructure. We would rather wikiprompt.org be the boring, reliable place everyone already has a copy of than the site people quietly route a scraper around. Openness was never the risk here. The walled garden was.