Camera moves for short AI B-roll: one small move, not a journey
A B-roll clip in a reel runs for three to five seconds, which leaves room for very little camera. Seven takes from one reel show what happened when my prompts asked for too much, what made the cut, how far the camera really moved, and every prompt. There is a camera skill to download at the end.
Published 23 October 20267 video takes compared6 min read
The reel is me talking on a beach about DIY jet kits for furniture, cut with AI B-roll: an armchair lifts off on four mini turbines, and a friend and I fly our sofa to a shop window for snacks. The spoken lines run from 2.4 to 5.9 seconds, and each clip has one line to cover.
All the camera failures below came from my prompts. The models did what I wrote, so the useful part is what to write instead.
A camera locked to the subject
For the flight home from the shop I wanted both faces to stay large, so the MiniMax H3 prompt said: “The camera flies backward in front of the sofa at the same speed, keeping the sofa centred and the same size in frame.”
H3 did exactly that. The sofa sits in the middle of the frame for the whole flight while the town slides past behind it, and it looks like a treadmill. A snorricam, a camera rigged to the actor’s body, gets the same locked look on purpose, for panic or disorientation.
On a journey the subject needs motion on screen: it enters, crosses, shrinks, leaves. Give the camera a speed relative to the subject, slower so the subject pulls away, and count what passes: “three rooftops and a small square sweep under the lens”.
A big move in a short shot
For a horizontal version of the B-roll I wanted a moving camera, and I wrote moves that would suit a 20-second shot. The football shot had to crane from the centre view up to look down over the sofa, then crane back, in 3.8 seconds. The prompt said “medium amplitude at slow speed”, but that path in that time cannot be slow.
MiniMax H3 Turbo · 4 s
Overhead in under two seconds
From eye level to directly above the sofa by 1.8 seconds, then back down through the lemon tree.
MiniMax H3 Turbo · 6 s
The world tips over and back
The chair lifts while the camera swings high above it and returns, all in about two and a half seconds.
A clip this short holds one camera move, the way a slider shot does on set. This is what I work to now:
Shot length
What fits
How to write it
3 to 4 s
One move at constant slow speed, in one direction
“a slow push-in, about 8% closer by the end”
5 to 6 s
One move, with a slow ease-out if you want one
“a slow lateral slide of half a metre”
8 s and up
A change of move can work
The relay below, for a journey between two places
One small move, with a number
I then made the same shots in Google Flow as 4-second clips, each with one slow move and a named size. Two of them are below. They look calm and the action reads, and these are the clips that went into the final reel.
Google Flow · 4 s
Button, jets, lift-off
The press, the ignition and the lift all read, because the camera leaves them alone.
Google Flow · 4 s
Turbines, a gentle drift
The exhaust keeps falling straight down and the background shifts a little.
Flow moved far less than the numbers I gave it. I matched the background in the last frame of each clip against its first frame, a rough check that is good to about 1%. The last two rows are the football and empty-snacks shots, which went into the reel too.
Shot
Asked for
Measured by the last frame
Armchair
About 5% closer
No closer; the frame shifted about 4% of its height
Turbines
15-degree arc, about 15% closer
No closer; about 2% of its width sideways
Football
About 8% closer
Back at the starting frame, after drifting about 3% wider midway
Empty snacks
About 8% closer
About 3% wider, the opposite direction
For B-roll under a voice that was fine, and I kept them. If you need a push-in people will notice, ask Flow for more than you want, then compare the first and last frames. H3 went the other way when I gave it room, so check each model separately. I have only these four Flow takes, and I have not run the same numbers on H3.
Pinned frames outrank the prompt, so check that the camera you describe can connect them. My approved armchair end frame was shot from a higher camera, which forces a big move however small you ask for it, so I wrote the Flow armchair from the first frame alone. When both frames share one framing, a move has nowhere to go. Cropping the last frame slightly tighter should give a push-in somewhere to land, but I have no take that proves it.
A journey between two places
The flight from the shop window to the home terrace is the one shot that needed a changing camera, because the first and last frames are in different places. It worked as one continuous take in four beats:
The camera stays by the window and the sofa flies out of the frame at the top right. Hold a moment on the empty sky.
The camera travels on its own, an FPV or drone flight over the rooftops.
The camera reaches the terrace first and settles on the empty mat.
The sofa drops into the frame from the top and slows into the last frame.
Leaving the frame and coming back gives the viewer the distance, and the faces only need detail in the first and last beats, where the pinned frames supply it.
MiniMax H3 Turbo · 5 s
All four beats in 5 seconds
The flaw is the drop-in: at about 4 seconds the sofa arrives as a smear and both faces melt for about four frames, so check that moment frame by frame.
Google Flow · 8 s
The take that went into the reel
Eight seconds gave each beat room, and I still had to cut it. The speech left 4.35 seconds, so I removed the middle of the flight and played the rest at about 1.1x. Asking for eight seconds was my planning mistake; a relay should start from the length of the line.
How to write the camera
Give the camera its own sentence, with the move, its size and its speed. Mixed into the subject’s sentence, the model may split the motion between them. The action goes in the first few seconds, because a 4-second clip may only get 2.5 seconds in the edit:
Camera: one slow, steady [push-in / arc / slide] for the whole shot, about [8% closer / 15 degrees / half a metre] by the end, constant slow speed, same height and angle.
Action, in the first three seconds: [what the subject does].
Describe what is in the frame and leave out “no X”. MiniMax H3 has no negative channel and tends to show what you name, so I wrote “the carpet stays flat on the deck” and left out “the carpet does not move”.
More moves, and the skill
For more moves, aicameramovements.com has 46 named camera moves with looping previews and a prompt to copy for each. Its prompts leave the size open, so add the number yourself.
I turned these rules into a skill that an AI agent can follow when it writes video prompts: move size by shot length, the four-beat relay, what to do about pinned frames, and which move fits which beat. Download the file and ask your AI agent to install this skill.
AI Video Production Lab is a live workshop for people already generating AI video who want control over consistency, references, camera movement, physics and failed generations.