How people actually prompt generative AI

1,000+ real prompts from viral posts on X — classified by model, technique, theme, and reference type. Every entry traces back to a practitioner who shipped something worth sharing.

A prompt observatory. Zero synthetic data. Zero crowdworkers. Just the organic distribution of how people talk to image and video models.

-
Prompts
-
AI models
-
Use references
0%
Synthetic

How it's built

No crowdworkers. No GPT-generated prompts. Every entry traces back to a real post by a real practitioner.

01

Collect

Viral image and video generation posts are sourced from X. Only posts with high engagement pass the filter — likes, reposts, and replies act as a quality signal.

02

Extract

Claude strips social framing ("check out this image I made!") to isolate the raw generation prompt. The model, technique, and reference type are detected automatically.

03

Label

Each prompt is classified across 5 dimensions: category, visual theme, art style, reference requirements, and model family. Every label links back to its source post.

Image vs. Video

The split between image generation and video generation prompts in the dataset.

Random prompt

Related prompt datasets

Several prompt datasets exist for evaluating generative AI models. This collection serves a different purpose — documenting organic practitioner behavior rather than providing evaluation prompts.

DatasetSizeSourceModalityEngagementCurated
DrawBench200Synthetic (LLM)ImageNone2022
PartiPrompts1,632Crowdworkers (Google)ImageNone2022
T2I-CompBench6,000Synthetic (GPT-4)ImageNone2023
GenAI-Bench1,200LLM + human mixImage + VideoNone2024
EvalCrafter700LLM + real usersVideoNone2024
VBench1,600Manual per dimensionVideoNone2024
T2VEval-Bench1,783LLM + manualVideoLab MOS2025
ummerr/prompts(this dataset)-Organic / in-the-wildImage + VideoViral filterMar 2026

Start exploring

Filter by model, technique, or theme. Copy what works. See what the community is actually making.

Browse prompts →