Using and Citing the Wikiprompt Corpus for Research

Using and Citing the Wikiprompt Corpus for Research

Prompt engineering produces text at scale, and text at scale is exactly the kind of thing researchers, students, and journalists like to study. What phrasing patterns recur across image prompts targeting different models? How do creative prompts differ structurally from coding or productivity prompts? These are answerable questions, but only if there is a corpus to answer them against, and until now that corpus lived scattered across the timelines of individual social accounts, un-indexed and effectively unstudiable.

Wikiprompt has spent the last several months curating exactly this kind of collection: public, user-submitted AI prompts pulled from the open web, organized and tagged. The catalog now holds over 55,000 prompts spanning text models like ChatGPT, Claude, and Gemini, and image or video models like Midjourney, GPT Image, Seedance, Veo, Kling, and Nano Banana. That corpus is available as a public bulk dataset, and this post is a guide to using it responsibly in academic or journalistic work: how to pull it reproducibly, how to attribute it, and where its limits are.

Why this corpus

Most prompt datasets that circulate in research contexts are small hand-collected samples or scraped snapshots with unclear provenance. Wikiprompt's differs in three ways.

First, every record traces to a real, attributable origin: an original_source field links to the original post, and an author field names the person who wrote it. This is a curated index of real public prompt-sharing activity, not a synthetic or anonymized dataset.

Second, it is structured and comparable across records. Every prompt has a category (creative, marketing, personal, productivity, coding, education, business, research, or other), tags, a model field naming the target AI system, and, where applicable, a metadata object with media_type, aspect_ratio, style, and an editorial quality assessment, unusual for a public dataset and useful for studying what separates a strong prompt from a weak one.

Third, it is accessible without friction: no API key, no rate limits to work around, no scraping infrastructure to maintain. That matters for reproducibility, which is the point of this post.

Pulling the data reproducibly

The dataset is exposed as two endpoints. The dataset manifest returns a JSON document describing the collection: total_prompts, the full record_fields schema, and the pagination scheme. The catalog itself lives at /dataset/prompts, also JSON, paginated with a keyset cursor rather than page numbers.

Keyset pagination is worth calling out because it is what makes a research pull reproducible. Offset-based pagination shifts under you if records are added or removed mid-crawl, silently duplicating or skipping rows. A keyset cursor does not have that failure mode: follow the next URL in each response until it returns null, and each page is anchored to a stable position rather than a row count that can drift.

A minimal pull:

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

limit accepts up to 500 records per page (200 by default), and the cursor parameter is after. A full crawl in Python is a short loop:

import requests

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

records = []

while url:

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

records.extend(resp["prompts"])

url = resp.get("next")

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

The response is served with Access-Control-Allow-Origin: * and heavy edge caching, so this loop is cheap and does not require a backend proxy. For a fixed, reproducible sample, record the date of your pull alongside your dataset snapshot, since the corpus is actively growing.

What is in a record

Each record includes fields for text analysis (title, description, content, the actual prompt text), for classification (category, tags, model), and for multimodal work (media, an array of image or video URLs, plus the structured metadata object). Timestamps (created_at, updated_at) support longitudinal study, and slug plus url give every record a stable, dereferenceable identifier for citing specific examples rather than only aggregate statistics.

To make this concrete, a citation might point to an architectural photography prompt built around a single brick casting a monumental shadow, or a stylistically distinct entry like a portrait combining autumn imagery with a Tibetan mandala motif, or one drawing on wuxia and thangka aesthetic conventions. Each of those pages is the canonical, citable location for that record: stable URL, visible attribution to the original author, and a link back to the original post.

For live filtering rather than a bulk pull, the search API supports query-based lookups over the same catalog, and machine-readable summaries live at llms.txt. Neither replaces the bulk dataset for systematic sampling, but both help with spot-checks or scoped samples, such as pulling only the creative category for a study limited to that domain.

Attribution and citation

Wikiprompt is an aggregator, not the original author of the prompts in the corpus. The content comes from public posts by individual creators, and correct attribution requires citing both layers: Wikiprompt as the dataset source, and the specific original_source for any record you quote, reproduce, or analyze in detail. We do not hold or claim a formal license over the underlying content; treat reuse as a request to attribute properly rather than a grant of broader rights. If a use case raises a question the dataset's fields do not answer, trace the record back to its original_source and note that lineage in your documentation.

A reasonable citation format for the corpus as a whole:

Wikiprompt Corpus. Retrieved [date] from https://www.wikiprompt.org/dataset/prompts.

Aggregated public AI prompt data; individual records attribute original authors

via the original_source field.

When citing an individual prompt, cite its canonical URL (wikiprompt.org/<slug>) and, where the analysis depends on original context, the linked original_source as well.

Limits worth stating plainly

This is a curated corpus, not a random sample of all AI prompt activity online. Only active, clean prompts are exported; anything removed for quality or policy reasons will not appear. That makes it well suited to studying what a moderated, public-facing collection of prompts looks like, but claims about "how people prompt AI models" in general should be scoped accordingly.

The corpus is also growing rather than fixed. A pull from one week will not exactly match a pull from the next. Where exact reproducibility matters, snapshot the data yourself (the keyset pagination above makes this cheap) and cite the retrieval date and, ideally, a record count of your snapshot, rather than pointing readers to the live endpoint and assuming it will match later.

Finally, the quality assessments in metadata are editorial judgments made during curation, not a formal or peer-reviewed rubric. They are useful as a variable to study, less useful as ground truth without disclosing that provenance.

None of this is a reason to avoid the corpus. It is the condition under which any aggregated public dataset should be used: know where the data came from, disclose the boundaries of the sample, and cite the people who actually wrote the prompts.