How to Write Camera Moves into Video Prompts

Use a vocabulary of real camera terms to direct motion in Runway, Kling, and Veo precisely.

7 minIntermediate

Visit Runway

The fastest way to get cinematic results from any video model is to speak its language, which is the language of real cinematography. These tools were trained on labeled film footage, so terms like dolly, pan, and crane produce reliable, predictable motion. This guide gives you the working vocabulary and how to combine it.

What you need

Any text-to-video or image-to-video tool
A scene you want to shoot
Willingness to test one camera term at a time
About 6 minutes

Step 1: Learn the core moves

Five terms cover most shots. Use them by name at the start of your prompt, where the model weights them most heavily.

Term	What the camera does
Dolly in / out	Moves physically toward or away from the subject
Pan left / right	Rotates horizontally from a fixed point
Tilt up / down	Rotates vertically from a fixed point
Crane / boom	Rises or lowers vertically through space
Tracking shot	Follows a moving subject alongside it

Step 2: Put the camera move first

Lead with the camera instruction, then the subject, then the lighting. The model treats the opening words as the strongest signal, so the move you name first is the move you are most likely to get.

camera-prompt.txt

slow crane shot rising up over a marble staircase,
a figure in a red coat ascending,
dramatic side lighting, cinematic.

Prompt structure

[ CAMERA ] slow crane shot rising up

[ SUBJECT ] a figure in a red coat ascending a staircase

[ LIGHT ] dramatic side lighting, cinematic

------------------------------------------------------------

-> model weights the first clause most

Camera first, subject second, lighting last.

Step 3: Add speed and intensity

Qualify the move with a speed word. Slow, gentle, and gradual keep motion controlled; fast, rapid, and whip add energy but raise the risk of distortion. Start slow, then increase only if the shot needs it.

One move per clip

Combining a pan and a dolly and a tilt in one short clip confuses the model. Pick a single dominant move per generation; chain different moves across separate clips in the edit instead.

Step 4: Test terms in isolation

When a move does not land, change only the camera phrase and regenerate with everything else identical. This tells you whether the model understood the term or whether the rest of the prompt was overriding it.

A/B camera test

Test A: 'slow dolly in' -> camera moves forward (good)

Test B: 'slow zoom in' -> lens zooms, flatter (different)

------------------------------------------------------------

scene + lighting held constant between the two

Same scene, two camera phrases, to compare motion.

Step 5: Build a shot list

For a sequence, write each shot as one camera move plus subject. A short list of single-move shots edits together into something that feels directed, because every clip has a clear, intentional motion.

Result

Your prompts now read like shot descriptions, and the camera does what you ask far more often. The simple habit of camera term first, one move per clip, is the difference between random motion and direction.