Put a second cat in the frame and the usual failures of AI cat videos change character: bodies fuse where they touch, a fifth leg appears at the seam, and the thing they were supposed to do with each other happens to no one in particular. This article makes two cats toss a cheese puff across a sofa, then three, then four, in CapCut at 16:9, five seconds a clip, and shows the writing that kept the bodies apart and the action readable. Six clips were generated on 8 September 2026. Five were downloaded and checked frame by frame; one additional two-cat clip could not be saved and is excluded from the results below.
Give Each Cat a Side and One Mark to Be Known By
The prompt for an interaction has to be able to say who does what, and it can only do that if the cats have names. They get them from position and from one visible feature each, fixed in a still before the first clip is asked for. The still costs one credit and settles the seating; the clips inherit it.
Photoreal still, 16:9 landscape, a bright living room in daytime. Two adult cats sit on a wide gray fabric sofa, facing each other with a clear gap of about one cat-length between them, both in full view from the side. On the left, a gray tabby cat with a white chest sits upright. On the right, an orange cat with a white-tipped tail sits upright. Between them on the sofa cushion is a small white bowl of round orange cheese puffs. Both cats look at the bowl. Camera at sofa height, straight on, eye level with the cats, no tilt. Soft window light from the left. No people, no text, no logos, no brand marks.
The seating plan. Left and right, one mark each, a gap you could name, and the object in the middle where both can reach it.
Three choices in that paragraph do the work later. The cats face each other from the side, so their outlines do not cross in the picture plane. The gap is described as a length, "about one cat-length", not as a feeling. And each cat has a mark the video prompt can point at without ambiguity: the tabby's white chest, the orange cat's white-tipped tail. The still came back through the CapCut AI image generator flow in Video Studio's Image mode at Seedream 4.5, 16:9 and 2K, exactly as seated.
One Sentence for the Action, One for the Object, One for Everyone Else
The video prompt is short and every sentence has a different job. One sentence names the actor and the action. One sentence follows the object on its own, start to finish. One sentence tells the receiver what to do. One sentence tells every cat to stay where it sits, and one asks for the anatomy to hold.
The camera stays fixed. The gray tabby cat on the left picks up one cheese puff from the bowl with its right front paw and tosses it in a short arc to the right. The cheese puff flies from the bowl to the orange cat. The orange cat on the right catches it in its mouth and chews. Both cats stay seated in their places, one on the left and one on the right, and never overlap. Each cat keeps four legs and one tail. Nothing else moves. No cut, no zoom, no pan.
The object sentence is the one people leave out. Without it, nothing in the prompt says where the puff starts or where it ends, and the two cats' sentences are left to imply it. Written as its own subject, "the cheese puff flies from the bowl to the orange cat", it has a start and an end that both cats' sentences agree with. The still was attached to the composer for the CapCut AI video generator in Video mode, with the settings panel set to Video clip, Seedance 2.0 Mini, 720p, 16:9 and 5s.
The Video mode settings used for every clip: 16:9 and 5s, Seedance 2.0 Mini at 720p.
The cost line the chat prints before each request: 1 credit for a still, 40 credits for a five-second clip at these settings, on 8 September 2026.
On 8 September 2026 the chat printed "Generating 1 item will consume 1 credit" before each still at Image mode with Seedream 4.5, 16:9 and 2K, and "Generating 1 item will consume 40 credits" before each clip at Video clip with Seedance 2.0 Mini, 16:9, 5s and 720p. Those lines are the only credit figures used here; other durations, models, resolutions and modes are priced differently, so read the line before you send. Six clips at those settings is the budget this article spent on video.
Two Cats, Three Clips, One Prompt
The same still and the same prompt were sent three times, because one good clip proves nothing about a method. All three came back with the action complete and the bodies apart. In each, the tabby reaches into the bowl at about one second, has the puff in its raised paw by two seconds, tosses at about two and a half, and the orange cat catches it in its mouth just after three and chews for the rest of the clip. Neither cat leaves its seat, the outlines do not touch, and in the frames checked at full size each cat has the legs and the tail it started with. The three clips differ in small things, the height of the arc, how far the orange cat leans, how long the chewing lasts, and agree on the sequence and its timing, which is what a method is supposed to buy: three of three passed the check described in the fifth section.
The first of the three two-cat clips, as downloaded: 1280 by 720, 24 frames per second, 5.04 seconds. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document.
The two-cat clip in nine frames. The action lives between 1 and 3.5 seconds; the rest is two cats sitting.
The puff's path, found by its color in every third frame from 1.5 to 4.4 seconds. It leaves the paw, climbs nearly half the frame and lands at the mouth, which is the sentence the prompt gave it.
The path drawing is the useful check for an object that passes between animals: it either has one start and one end or it does not. Here it does. One detail to know before you decide the clip is right for you: in all three clips the tabby holds the puff between the toes of a raised paw, the way a hand would, rather than batting it. That is how the model interpreted "picks up with its right front paw" in this run, and it reads as the cartoon version of a cat rather than the animal. Wording that asks for a nudge or a flick instead of a pick-up was not tested here.
One more two-cat clip was generated and is not in the count. It finished, but a "Couldn't save project" dialog covered the canvas before the file could be downloaded, and the tab was closed before that was noticed; its frames were not inspected, so it is neither a pass nor a failure. Download a clip the moment it appears.
Add a Third Cat, Then a Fourth
The same method was pushed to a row of three and then four, with the bowl moved in front of the cat that acts and an explicit sentence for the ones that do not. The three-cat prompt names the actor by position and mark, "the black cat in the middle", and the bystander by what it does, "the gray tabby on the left watches and does not move".
The camera stays fixed. The black cat in the middle picks up one cheese puff from the bowl with its right front paw and hands it to the orange cat on the right. The cheese puff moves from the bowl to the orange cat's mouth. The orange cat takes it and chews. The gray tabby on the left watches and does not move. All three cats stay seated in their places and never overlap. Each cat keeps four legs and one tail. Nothing else moves. No cut, no zoom, no pan.
Three cats: the black cat hands the puff to the orange cat while the tabby watches. A GIF version is available.
The handoff at full pixel size. The paw and the mouth meet at one point and separate again; nothing merges.
Four cats in a row, same method: the black cat feeds the orange cat, the tabby and the white cat stay put. A GIF version is available.
Four cats in nine frames. The two outer cats do what their sentence said: nothing.
Both came back complete. In the three-cat clip the black cat has the puff at 2.5 seconds, the orange cat takes it at 3.8 and the tabby does not move; averaged over the whole frame, the picture changes less from frame to frame than the two-cat clips do, because two of three bodies are still. In the four-cat clip the timing is nearly the same and the white cat on the far right, which was given the same "watches and does not move" instruction as the tabby, holds its pose. The row worked at four in this run, seated, side by side, at eye level, with one action between two neighbors. What the extra cats cost is attention: with four in the frame the puff is small and the handoff happens in a tenth of the picture, so the clip is better on a large screen than in a feed.
The same grip in the two-cat and four-cat clips: the puff held between the toes of a raised paw. Consistent, and not how a cat holds anything.
Count Paws Before You Count Anything Else in a Multi-Animal Clip
A multi-animal clip is not judged by watching it once. The check that caught nothing wrong here, and would have caught the usual failures, takes a minute per clip. Pull nine frames at equal intervals and look at the outlines: two cats should be two shapes in every frame, with sofa between them. Then take the frames where the paws are doing something, roughly one to three and a half seconds in these clips, and look at them at full pixel size, not scaled to a phone: count front paws on the acting cat, look at the seam where a paw crosses another body, and follow the object from one frame to the next to see that it moves rather than reappears. The tails are the last thing to check, because they are the first thing to double when two bodies sit close.
If you want a number to point at instead of an eye, the average change between consecutive frames tells you where the action is: in these clips it is near zero except for the window in which a paw moves, which is where to look. The object path drawn above is the same idea for the thing being passed. Neither replaces the paw count; they tell you which frames to count in.
Where This Stops
Everything above holds for a row of seated animals, side by side, at eye level, doing one thing between two neighbors with one small object. That covers a large share of the multi-animal clips people actually want, and in this run it covered two, three and four cats without a failed clip among the five inspected. It says nothing about the harder cases, which were not tested: animals that stand and cross each other's path, two animals touching along their whole length as in a hug, more than four in the frame, a moving camera, or an action faster than a toss. Each of those removes one of the things the sofa row gave for free, a fixed position, a visible gap, or a single seam.
Use this structure as a starting prompt pattern, not a guarantee. Generate more than one version when the result matters, then keep only clips that pass a frame-by-frame anatomy and interaction review.
Written 8 September 2026. All cats, stills and clips are generated images and video from CapCut made for this article; no real animals, products or brands appear. Interface labels, panel contents, credit lines and clip specifications reflect a single session on that date and the configurations named above, and may change. Frame timings are read from the downloaded files at 24 frames per second; the object path was located by color in the frames.