About vflow
vflow is a structured index of AI video prompts. Every prompt is a public post written by someone else; what we add is the analysis layer on top — the camera, lighting, duration and constraints that make a prompt reusable rather than just readable.
What we add, and what we do not
| Prompt text | Reproduced verbatim from the original public post. Never edited, never translated, never summarised. |
|---|---|
| Video | Never copied or re-hosted. Playback streams from X's own servers; the original post is always one click away. |
| Attribution | Every page names the author and links to their profile. |
| Our addition | A structured breakdown we produce ourselves, plus the aggregate statistics on the statistics page. |
| Model training | We do not use this content to train or fine-tune models. |
How the prompts are analysed
This is the part worth being precise about, because it determines what the numbers on this site do and do not mean. Each of the 5,926 prompts was processed in two stages.
1. Deterministic extraction
Values with an unambiguous written form — aspect ratio, stated duration — are pulled out with pattern matching. This stage is precise but conservative: it only fires when the prompt states the value explicitly, so it finds fewer results and almost no false ones.
2. Model reading
Everything requiring judgement — the title, the summary, the subject, the mood, and the camera and lighting terms — is assigned by a language model reading the full prompt text. Each field must come from a fixed controlled vocabulary; free text is rejected rather than stored. That constraint is what keeps the vocabulary translatable and the filters consistent, and it means a term being absent is a statement about the prompt, not a guess by us.
What that means for the numbers
Coverage figures are shares of prompts that state a field, not claims about what the videos contain. Camera language appears in 89.5% of prompts and lighting in 83%; the remainder simply do not say. A prompt not mentioning lighting is not a prompt without lighting.
Where this data is weak
We would rather state these than have you discover them:
- The corpus reflects who posts, not what works. It shows what creators write, not which prompts produced good video. Nothing here is a quality ranking.
- It is a snapshot. Source posts run from 2025-10-01 to 2026-09-10. Prompt practice moves quickly, so treat older entries as history.
- Duration is the least-stated field (52.6%), so duration statistics rest on a smaller sample than the others.
- Language-model fields can be wrong. We validate vocabulary and reject out-of-vocabulary values, but a model can still misread an unusual prompt. Corrections are welcome.
Corrections and removal
If a page misrepresents your work, or you want it gone, write to REPLACE_WITH_REAL_EMAIL. Removal requests need no justification and are actioned within 48 hours — see removal requests and information for authors. Corrections are handled the same way: we would rather fix a page than lose it.
How this site is built
A static site — plain HTML, no account, no tracking cookies, and no JavaScript required to read any page. Filtering and playback are enhancements layered on top of HTML that already works without them. Statistics on this site are computed from the same corpus that generates the pages, so they can be reproduced rather than taken on trust.