Explained · updated 2026-10-14
The Crazy Frog AI prompt
There is no single magic Crazy Frog AI prompt. The videos in the trend are made with reference-to-video models: you give the model the original chase as a reference clip, a photo of the new rider, and a prompt that says exactly what to keep and what to change. This page explains what that prompt needs, shows a full example we wrote, and explains why short prompts usually fail.
Short answer: the prompt describes the invisible moped pose in detail, each shot in order, who the rider and chaser are, and where it happens. The reference clip carries the moves and the timing.
The prompt the trend started with
On October 8, 2026, TikTok creator @tonnatusha, who posted one of the first AI versions the day before, shared the prompt and instructions in Russian. The post says they are long, so they were shared as a file on Yandex Disk rather than in the caption. We did not open that file, so we do not quote it here. What follows is our own prompt for the same scene, a simplified version of the one our maker sends to the model.
What a Crazy Frog AI prompt needs
- The reference rule. Tell the model to take motion, camera and timing from the reference clip only. Without this, it copies the original rider's face as well.
- The invisible moped pose, spelled out. Low squat, feet off the ground, fists on handlebars that are not there, and a plain statement that there is no moped. Models love to add a real scooter.
- Each shot, in order, with its time. The rider, the chaser close-up, the flip, the getaway. One sentence per shot keeps the cuts where they belong.
- Each person, separately. Who the rider is, from which photo, and what to keep (face, hair, clothes). The same for the chaser, or a description of one if you have no photo.
- The set. The place, its light and weather, and "no other people", so the model does not fill the road with extras.
- The beat. If your tool accepts an audio reference, ask it to keep the moves on the beat of that audio.
An example Crazy Frog AI prompt
This is the structure our maker uses, simplified. Image 1 is the rider photo, Image 2 is the background, Video 1 is the reference chase and Audio 1 is its sound.
Use Video 1 for motion, camera and timing only. Copy every cut, camera move and body pose exactly.
PERSON 1 (the rider, from Image 1): keep their face, hair, skin tone, build and clothes.
In every shot they squat low in mid-air as if sitting on a small moped that is invisible:
knees bent, feet off the ground, both fists held forward on invisible handlebars.
There is no moped, no bike and no seat under them.
Shot 1 (0-5 s): the rider glides toward the camera along the road, grinning, leaning into the turns.
Shot 2 (5-7 s): close-up of the chaser on a white futuristic hover-bike with round red lights.
Shot 3 (7-11 s): the chase; the rider flips over in mid-air and lands still squatting.
Shot 4 (11-14 s): wide shots of both racing away into the distance.
PERSON 2 (the chaser): a motorcycle police officer in a dark uniform and white cap,
riding the hover-bike behind the rider.
SET (from Image 2): the empty road in this picture, with its light and weather.
No other people. Keep the timing of every move on the beat of Audio 1.Our prompt, not the trend creator's. Results vary by model and by attempt.
Why one-line prompts fail
"Make me ride an invisible motorcycle like the Crazy Frog trend" gives the model almost nothing to hold on to. Common results are a real scooter appearing under the rider, the original person's face staying in, the flip happening at the wrong moment, or extra people on the road. Being specific about pose, people, shots and set fixes most of these. Even so, general tools often need several tries, and each try costs credits.
Which tools take this kind of prompt
Reference-to-video models that accept a video plus images, such as the MiniMax H3 family our maker uses, or tools built on similar models such as Higgsfield Genjutsu. Text-only video models cannot copy the exact moves of the trend clip, so the result is a different video. Video models also refuse photos of public figures and often refuse copyrighted characters.
Skip the prompt
If you just want the video, our Crazy Frog AI video maker has the reference clip, the pose, the shots and the 16 backgrounds built in. You add a photo, optionally a chaser, pick a place and press one button. The how-to guide covers photo tips, and the trend page covers where it all came from.