AI Image & Video

Camera Movement Prompts for AI Video: Stop the Locked-Off Shot

I went back through my old prompts: 29 of 32 shots said "eye-level" and 29 said "static camera". The AI wasn't ignoring me; I never told it how the camera should move. In five minutes this note shows you how to rewrite any AI video prompt so the camera moves, and moves for a reason.

10/26/2026 ｜ 2 min read

What This Note Gives You

In five minutes you can rewrite any AI video prompt so the camera moves and moves well. It works for typical image-to-video and text-to-video tools.

One Core Idea

When an AI video camera never moves, the model usually isn’t disobeying. The prompt never said how the camera moves, or it said “static camera”. In my own old prompts, 29 of 32 shots said “eye-level” and 29 said “static camera”.

How To Do It

Step 1 | Count first

Pull every prompt from your latest video and count how many shots say “eye-level”, “static”, “locked-off” or similar.

Step 2 | Write three things per shot

Where the camera goes (push in, pull out, orbit the subject, rise), how fast (slowly, within one second), and where it stops (ending on a close-up of her face). “Push in” alone often stops at an odd spot.

Step 3 | Tie the move to a story beat

For example: “the moment he grabs her wrist, the camera rushes in to a close-up of the wrist within one second.” A motivated move doesn’t feel random.

Step 4 | Pick the stronger version of the same idea

To get closer to a character, try “orbit her while closing in” before “slow push in”. In my blind test of 30 camera moves, all 4 orbit moves were rated good, while a slow push-in only rated average.

Step 5 | No fixed quota; follow the scene

Tense scenes can take more aggressive moves. In tender scenes, or a beat where the line must land, slow down or hold still.

Validation

Same scene, three versions: the aggressive-move version rated “good”, the restrained one “average”, and the nearly static one “bad” no matter how pretty the frame. On 3 unseen scenes, the new method went 6 for 6 “good”, averaging 2.00, versus 1.83 for my old approach and 1.50 with no method.

Pitfalls You Will Hit

1. Testing with the wrong material: my first round was 79 still images, and you can’t see camera movement in a still. To judge camera work, watch real video.
2. Writing moves AI can’t render: the dolly zoom (subject stays the same size while the background stretches away) failed 3 out of 3 tries. Lateral tracking, heavy shake and frozen-time orbits also fail often. Swap in an orbit or a slow push-in.
3. Two moves in one shot: “rise while pulling back” often drifts into a different scene. One shot, one move.

When You Get Stuck, Ask an AI

Paste this into Claude or ChatGPT with your own prompt:

I'm a beginner generating AI video. Below is the prompt for one shot, and the camera doesn't move at all.
Rewrite it so the camera moves. Always state: where the camera goes, how fast, and the frame it ends on.
One move per shot, and avoid moves AI often can't render, like dolly zooms and lateral tracking.
My prompt: <paste here>

One-Line Summary

For every shot, say where the camera goes, how fast, and where it stops; pick the stronger move and adjust to the scene. To judge it, watch real video.