Some videos hold attention without a novel subject or a clever story. What makes them work is how the story is shown: the camera angle, the quality of light, the order of the scenes. That craft is cinematography, and with the arrival of AI video generators it is no longer confined to directors with large budgets. Anyone prompting a tool such as OpenAI’s Sora, Runway or Pika is now, in effect, directing a camera.
The catch is that these tools do not choose a shot for you. They follow the description they are given. A prompt that says only “a girl standing on a beach” leaves framing, distance and mood to chance. A prompt that specifies a wide shot at sunset with the figure alone in the frame tells the model both what to show and how to show it. OpenAI’s own Sora 2 prompting guide makes the same point, noting that a wide shot from above emphasizes space and context while an eye-level close-up focuses attention on emotion.
The basic shot vocabulary
Filmmakers have a shared language for framing, and each term carries an intended effect. The most useful ones for AI prompting are these.
- Wide shot: shows the full environment, such as a city, a desert or a room. It establishes context and is typically used to open a scene or introduce a location.
- Close-up: frames a face or a single object tightly. It highlights emotion, draws attention to detail and signals that a moment matters.
- Medium shot: shows a character from roughly the waist up. It is the default for dialogue and gestures because it balances expression with body language.
- Over-the-shoulder shot: views one person from behind the shoulder of another, usually during conversation. It conveys relationship, tension and point of view.
- Low angle: places the camera below the subject looking up. It makes a character appear powerful, imposing or threatening.
- High angle: places the camera above the subject looking down. It makes a character appear small, vulnerable or overwhelmed.
Why it matters more with AI
In traditional production the cinematographer makes these choices on set. With generative tools the prompt is the only set. Runway’s Gen-4 prompting guide recommends describing camera position and movement explicitly, and OpenAI’s guidance for Sora 2 lists framing instructions such as “wide establishing shot” and “aerial wide shot” as examples of effective direction. Without them, models fall back on generic framing, and the result tends to feel flat.
Consider a prompt for “a boy sitting under a tree looking at the sky.” The scene could be rendered as a wide shot that emphasizes the tree and the open sky, suggesting solitude and scale, or as a close-up on the boy’s face that foregrounds his expression and state of mind. The same subject produces two different emotional outcomes depending on one phrase.
Writing shots into a prompt
A workable pattern is to state the shot type first, then the subject and action, then the lighting and atmosphere. For example: “Wide shot, a woman stands alone on a beach at sunset, warm golden light, gentle waves.” For a dialogue scene: “Over-the-shoulder shot, two colleagues argue across a desk in a dim office, cool fluorescent light.” For a moment of menace: “Low angle, a man in a long coat walks toward the camera down a rain-soaked street at night.”
Combining shot type with camera movement adds another layer. “Slow push-in on a close-up” or “tracking shot following the subject from the side” are both widely supported phrasings, though results vary by tool and by model version.
The tool follows, the creator leads
AI video generation does not replace creative judgment. It executes it quickly. A creator who knows the language of cinema can turn a vision into footage in minutes; one who relies on generic descriptions will get generic clips. Whether the goal is a marketing video, a short film or a social media clip, understanding what each shot communicates remains the most reliable way to get results that viewers feel rather than merely watch.
