Skip to main content
·5 min read

Why Your AI Video Prompts Look Like Stock Footage (and the Camera Language That Fixes It)

Every AI video model has a default camera, and it's the same one: a slow, smooth, slightly floating move that drifts toward the subject. It looks expensive for two seconds and generic by the third.

That's the stock footage feeling. Nothing is wrong in the frame. The camera just has no reason to move.

On a real shoot, every move has a motivation. The camera follows someone, reveals something, or reacts to something. When the prompt gives the model that motivation, the clip stops floating and starts telling you where to look.

Describe the move, then the reason

The structure we use for the camera line:

[shot size] [move] [direction/speed], [what the move reveals or follows], [lens feel]

Example:

Medium shot, slow push-in toward her face as she hears the door open behind her, shallow depth of field, 50mm feel

The push-in alone is a technique. "As she hears the door open" is a reason. Models handle the second version better because the move now has a beat to land on.

The moves worth knowing

You don't need a film school glossary. These cover most of what you'll want:

  • Static / locked-off — the camera doesn't move. Underrated. It makes action inside the frame feel more intense.
  • Push-in (dolly in) — the camera moves toward the subject. Builds tension or intimacy.
  • Pull-out (dolly out) — moves away. Reveals context, isolation, scale.
  • Tracking / follow — moves alongside or behind a moving subject.
  • Pan — rotates left or right from a fixed point. Following action, scanning a space.
  • Tilt — rotates up or down. Revealing height.
  • Orbit — circles the subject. Hero moments, product shots. Easy to overuse.
  • Crane / jib up — rises vertically. Openings and endings.
  • Handheld — slight organic shake. Urgency, documentary feel.

Pick one per shot. Asking for a push-in, then an orbit, then a crane up in a five-second clip is how you get the warping, melting camera that makes AI video obvious.

Shot size sets the scale of the move

The same move means different things at different sizes.

A push-in on a wide shot is an approach. A push-in on a close-up is a thought. Always say the shot size: extreme wide, wide, medium, medium close-up, close-up, extreme close-up.

Speed is a mood

"Slow" is the model's default. Saying it changes little. Saying the opposite changes a lot.

  • "Slow, almost imperceptible push-in" — dread.
  • "Quick whip pan to the right" — surprise, energy.
  • "Steady tracking shot at walking pace" — grounded, observational.

If the scene has a moment, put the speed change on the moment: "static, then a sudden push-in as the glass breaks."

Lens feel instead of lens specs

Millimeter numbers work as shorthand for a look, not as physics. Models have absorbed what "24mm" and "85mm" images look like.

  • Wide (16–24mm feel) — space, distortion near the edges, energy.
  • Normal (35–50mm feel) — closest to the eye, honest.
  • Long (85mm and up) — compressed background, isolation, intimacy.

Pair it with depth of field: "shallow depth of field, background softly out of focus" or "deep focus, everything sharp."

Four prompts, four motivations

Follow

Wide tracking shot following a cyclist from behind through a narrow market street at dusk, camera at handlebar height, steady, warm string lights overhead, 24mm feel

Reveal

Close-up on a pair of muddy boots at the edge of a cliff, then a slow tilt up to reveal the hiker looking over a sea of clouds, soft morning backlight

Tension

Locked-off medium shot of a man at a kitchen table, phone face down in front of him, he doesn't move, the only motion is the steam from his coffee, low-key light, long hold

Energy

Handheld medium shot running alongside a dancer in an empty parking garage, quick reframes, hard overhead fluorescent light, 35mm feel

Notice the third one barely moves. Sometimes the strongest camera direction is "don't."

Model-specific notes

Different models respond to camera language with different reliability, and they change with every version. A few patterns that hold up:

  • Put the camera line early. Some models weigh the beginning of the prompt more.
  • Keep the subject's action and the camera's action separate in the sentence, so the model doesn't confuse "she turns" with "the camera turns."
  • If a model supports explicit camera controls or start and end frames, use them for the move and keep the text for the motivation.

When a model gets a move right, that exact phrasing is worth more than any guide, including this one. Write it down.

Build a shot library, not a prompt list

The moment you get a clean push-in on one model, you'll want it again on a different subject, maybe a different model. The text alone won't tell you which version worked. The clip will.

Keep every shot that works as a card: the prompt, the model it ran on, and the result next to it. That's the whole idea behind Super Prompts — a library where every prompt comes with its proof.

Lighting is the other half of the frame. If you haven't already, read how to write cinematic lighting prompts.

Save these prompts in one place

Super Prompts lets you organize, search, and reuse AI prompts across 25+ tools. Free to start.

Try Super Prompts Free