How do you write prompts that actually get usable AI video and images?

Treat every prompt like a director's brief: subject, action, scene, camera, lighting, and audio, one shot per prompt, fully re-described every time because the model has no memory between shots. The official Veo and Sora guides plus the JSON-prompting pattern (structured fields instead of prose) are the fastest way in, and reference images beat words for locking a look. Expect roughly a 20:1 generation-to-keeper ratio and plan your iterations, not a single perfect prompt.