Every prompt lists the camera move, lighting and duration behind its clip, read from the text, not guessed. Filter by technique, not just by model.
A person recording a smartphone street vlog discovers that touching everyday urban objects instantly transforms them into magical versions while maintaining their original position and form.
Read the prompt →Four things hold across this set, each measured against the 376 prompts above rather than asserted.
Same measurement, applied per model. A lower number means that model's writers leave more to the model, not that the prompts are worse.
| Library | Prompts | Names camera | States duration | Locks consistency |
|---|---|---|---|---|
| Seedance 2.5 | 270 | 98.5% | 100% | 67% |
| Seedance 2.0 | 56 | 100% | 100% | 57.1% |
| Grok Imagine | 2 | 100% | 100% | 100% |
| MiniMax H3 | 32 | 100% | 100% | 56.3% |
| Veo 3 | 1 | 100% | 100% | 0% |
| Kling | 15 | 100% | 100% | 40% |
| Muse AI | 0 | 0% | 0% | 0% |
Promptweet is a prompt library: real AI-video prompts pulled from posts on X, each paired with the actual clip it produced and a spec sheet read from its own text. Most clips here are real generations, linked back to their source post. A few library corners are still filling in and show placeholder content in the meantime, always marked as sample data. Browse by model, filter by camera move or lighting, open a prompt, copy it.