The format is an office interview. Someone talks to an off-camera interviewer about something ordinary, and behind them, in the same shot, something absurd goes on without comment. The whole joke sits on the word: nobody acknowledges it, the camera does not swing over to look, and the audience is left to spot it unaided.
The standard advice is to keep the performance flat. That advice is correct and it is also incomplete, because the sentence you write about performance does not only reach the performer. Three clips were generated to find out what else it touches.
The scene, identical in all three. An office worker in a plain shirt sits in a swivel chair facing slightly off-camera, being interviewed, in a plain open-plan office. Behind him, a person in a full-body cartoon dinosaur costume calmly waters a row of desk plants with a small watering can. Handheld documentary camera, slight natural shake, interview framing from chest up with the subject placed off-center. No text, no logos, no printed words and no brand marks anywhere in the frame. 16:9 landscape.
Two clips added a flat-performance line. One added the opposite.
Flat. The seated man keeps a flat, matter-of-fact expression and does not react to the dinosaur at all, gesturing only slightly as he talks. The dinosaur goes about the watering calmly with no comic movement.
Broad. The seated man reacts to the dinosaur with big expressive gestures and exaggerated facial expressions, and the dinosaur moves in a broad comic way.
Measuring the camera and the actor separately
Two numbers were taken from every pair of consecutive frames. The first is how far the whole picture moved, found by phase correlation between the two frames — that is the camera. The second is how much of the picture still changed after that whole-frame movement was cancelled out — that is everything moving inside the shot, which here is the performer and the dinosaur. Splitting them matters because handheld camera work and a busy performance both read as "lots of motion" and they are not the same problem.
The broad clip has about 1.9 times the movement inside the frame, which is what was asked for. It also has about 1.9 to 2.8 times the camera movement, which was not — the camera wording was identical in all three. A quarter more of its frames are in motion rather than held.
The third thing that changed
The numbers say the camera moved. The frames say why, and it is the part that actually breaks the joke.
Same scene description, same camera description. The only difference is the line about how broadly to play it.
In the flat clips the framing is a front-on interview: the man talks toward the interviewer, the dinosaur goes about its watering behind him, and the two do not interact. In the broad clip the shot has become an over-the-shoulder reverse, and the man has turned around to look at the dinosaur and is gesturing at it.
That is a different scene. The premise of this format is a person who does not acknowledge the absurd thing; once the character turns to face it, there is no longer anything for the audience to notice on its own, and the camera has repositioned itself to watch the reaction rather than to record an interview. Asking for a big performance did not produce the same joke played louder. It produced a different and weaker structure.
So the flat-performance line is doing more work than it appears to. It is not only a note about the actor's face. It is what keeps the interviewee facing the interviewer, and therefore what keeps the camera where an interview camera belongs.
The flat version. 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds.
The handheld request does land
Worth noting separately, because it is the other half of the format and it behaved well. All three clips were asked for a handheld documentary camera with slight natural shake, and all three have real sub-pixel to one-pixel drift rather than a locked-off tripod look: mean movement between 0.18 and 0.52 pixels, with roughly three quarters of frames essentially held in the flat versions.
That is the right amount. Handheld in this format is not a visible wobble; it is the absence of the machine-perfect stillness that reads as a tripod. One pixel of drift at the 95th percentile is enough to remove the tripod feel without drawing attention to the operator. The wording that produced it is plain: handheld documentary camera, slight natural shake. Asking for more than that is the same trap as asking for more performance.
What to write, and in which order
Two of those three lines are about restraint and neither of them is about the joke. The absurd thing carries itself; the writing exists to stop everything else in the frame from reacting to it.
What it cost
75 credits. The figures are what the interface showed for this configuration on 16 September 2026; the panel states the cost before you send. One artifact worth reporting rather than hiding: in the flat takes the seated man's face smooths out and loses detail for part of the clip, which is a generation flaw and not something the wording caused. At the size these get watched it goes unnoticed, and at any larger size it would need a reroll.
Write the restraint before you write the joke
The instinct with a comedy prompt is to spend the words on the funny thing — the dinosaur, the costume, the absurdity — and to treat the rest of the description as setup to get through. The measurements here point the other way. The absurd element needed one clause and behaved. The two lines that decided whether the clip worked were both about things em>not/em> happening: the man not reacting, the background performer not playing to camera.
So draft in that order. Write the scene, then write down everything in it that must stay unimpressed, and only then write the thing that is meant to be funny. If you find yourself adding intensity anywhere — a bigger reaction, a wilder camera, a broader background gag — check what else it is likely to drag with it. In this test one sentence about performance also moved the camera, changed the shot type and turned the subject around, and none of those three were asked for.
Written 16 September 2026. All three clips were generated in CapCut for this article; no footage or photograph of a real person was used. Clips are 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds. Camera movement is the per-frame-pair global translation estimated by phase correlation on 640-pixel-wide grayscale copies, reported in pixels at that width; movement inside the frame is the mean absolute difference between consecutive frames after the estimated global translation has been cancelled and a 24-pixel border discarded, on a 0 to 1 intensity scale. The two are reported separately because handheld camera work and busy performance are indistinguishable in any single whole-frame motion figure. The flat condition was run twice and the broad condition once, so the flat figures carry a repeat and the broad figure does not; the gap between them is larger than the gap between the two flat takes, which is the only comparison the design supports. The framing and eyeline differences described above are visible in the published frames rather than measured, and are offered as observation rather than as a second measurement.