What 55,000 Prompts Reveal About How People Use AI

What 55,000 Prompts Reveal About How People Use AI

Every prompt on Wikiprompt started as a real request someone made to a real model, then decided was worth sharing. Multiply that by 55,000+ and you stop having a list of examples and start having a corpus: a rough census of how people actually talk to AI when they want something specific out of it. We built the dataset so anyone could pull that whole corpus and look for the patterns themselves, instead of trusting our summary of it.

This post is that summary anyway, but framed as an invitation. Everything below is a starting hypothesis, not a finding we're asking you to take on faith. The data is sitting at /dataset/prompts, it's free, and it takes about ten lines of code to pull.

What's actually in a record

Each entry in the dump carries slug, url, title, description, content (the prompt text itself), category, tags, media, model, a metadata object with structured fields like aspect ratio and style, author, original_source, and both timestamps. That's enough to run three different kinds of analysis without touching the live site: text analysis on content and title, categorical analysis on category/tags/model/metadata, and time-series analysis on created_at if you want to see how phrasing or model choice shifted over the life of the catalog.

The model landscape

Group records by the model field and you get something close to a leaderboard of what people are actually reaching for, not what's newest or most hyped. Image and video generators (Midjourney, GPT Image, Seedance, Veo, Kling, Nano Banana) sit alongside text and reasoning models (Claude, GPT, Gemini, Grok) in the same table, because Wikiprompt doesn't separate "chat prompts" from "image prompts" as different products. That's worth exploring on its own: does a given model attract a narrower or wider spread of categories? Do certain models cluster around certain styles in metadata? The metadata.model and metadata.style fields exist specifically so you don't have to regex the prompt text to answer that.

Image, video, and text aren't evenly split

metadata.media_type tags each record as image, video, or text, and the ratio between the three is itself a signal about where the interesting prompt engineering is happening right now. Text prompts (system prompts, personas, structured templates) tend to be longer and more scaffolded. Image prompts lean on dense visual vocabulary packed into one or two sentences. Video prompts, the newest and smallest slice, read more like shot lists, with camera movement, duration, and pacing spelled out explicitly. If you're trying to understand how "prompting" as a skill differs across modalities, filtering the dump by media_type before you do anything else will save you from averaging three different disciplines together.

Categories and styles, with real examples

The catalog groups everything into nine categories (creative, marketing, personal, productivity, coding, education, business, research, other), and creative is a good place to see style vocabulary at work. Take a minimalist black-and-white editorial smoking poster: the prompt is doing a lot of work with restraint, negative space, and a named photographic tradition rather than piling on adjectives. Compare that to a prehistoric creature sovereign portrait, which leans on genre-mashup language (regal, mythic, sovereign) to force a specific tone out of the model, or a pastel fantasy portrait of a young woman, which is almost entirely color and lighting description. Three prompts, three completely different strategies for getting specificity out of an image model, and all three are sitting in the same category: creative bucket. Tags and metadata.style are the fields that let you split "creative" into something more granular than the top-level category alone would show.

Phrasing patterns worth hunting for

A few things are worth testing against the full corpus rather than trusting anecdotally: whether certain opening constructions (imperative verbs, "you are a...", scene-setting clauses) correlate with quality/assessment scores in metadata; whether prompts targeting a specific named model use different vocabulary than model-agnostic ones; whether tag co-occurrence reveals categories that behave like each other even though they're labeled differently (marketing and business prompts, for instance, tend to share a lot of tags). None of this requires anything fancier than counting n-grams and grouping by field. It's the kind of dataset where a slow afternoon with pandas or a spreadsheet turns up something real.

Pulling it yourself

The manifest at the dataset tells you the total count, the record shape, and how pagination works before you fetch a single record. The actual data is keyset-paginated, 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. A minimal loop looks like this:

import requests

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

records = []

while url:

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

records.extend(res["records"])

url = res.get("next")

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

No API key, CORS enabled, and the whole thing sits behind an edge cache, so a full pull is fast and doesn't lean on our servers. If you'd rather query than bulk-download, the search API covers ad hoc lookups, and llms.txt documents everything machine-readable in one place.

Attribution, and what we'd like to see

Every record traces back to a real author through original_source, so any analysis, chart, or writeup that comes out of this data should credit both wikiprompt.org and the original creators it aggregates. We don't hold a formal license over the underlying prompts; we're the catalog, not the copyright holder, and the dataset is offered on that basis: free to explore, please attribute.

If you run something interesting on the corpus, from a model-adoption chart to a style taxonomy to a study of how prompt length correlates with quality scores, we want to see it. Fifty-five thousand prompts is enough signal to say something real about how people are actually using AI right now, and the honest answer is we haven't run every analysis on it ourselves either.