Crazy Frog AI prompt: what MiniMax H3 actually reads

This is the prompt the homepage maker sends to MiniMax H3 with your two people and the background. You do not type it in the tool. The photos decide who is on the moped; this text decides the crouch, the four shots, and what must never appear.

If you are using another generator, copy the block at the bottom. If you are using this site, skip to how to pick the photos and pick a background.

What each input is

The four shots are the format: wide glide, chaser close-up, mid-air flip, wide getaway. Timings for 10s vs 15s are on the video page. The presets that lock those beats are on the template page.

Shot list

Shot 1 — Wide. Subject 1 glides toward camera, grinning, leaning into the turns. The hover-bike follows.

Shot 2 — Close-up. Subject 2 grips the hover-bike, focused on the chase.

Shot 3 — Medium. One clean mid-air flip, land, back into the same crouch. No bike appears under them.

Shot 4 — Wide. Both race toward the vanishing point. Subject 2 is still chasing.

Keep the same road and direction of travel in every shot. Cut on the rhythm of the reference, not on a dissolve.

Words that break a clip

Leave these out of any rewrite, and out of Image 3:

• moped, motorcycle, scooter, seat under the rider — the joke is empty space under the crouch
• gray silhouettes / gray figures — motion reference only; never render them
• extra people, police car — two subjects, one hover-bike
• text, watermark — trips safety and looks like a stamp
• face swap between subjects — Image 1 stays the rider, Image 2 stays the chaser

Cinematic, 4K, masterpiece add noise. Names of real people do not reach the model and can trip filters. Use "Subject 1" and "Subject 2".

Image 3: the empty road

The background picker on the home page is Image 3. Halloween graveyard, overpass, neon city, desert, snow, beach, Mars, garden path, clouds, candy road, chocolate road, pixel brick road — or your own wide still.

The still should look like a road with nobody in it. Faces, cars, and shop signs in the background are how you get extra people or a real vehicle in the chase.

The prompt the maker sends

subject_definitions:
<Subject 1> is the rider, based on <Image 1>.
<Subject 2> is the chaser, based on <Image 2>.
<Image 3> is the empty road setting only. It contains no people or vehicles.
<Video 1> is the reference for shot order, camera movement, pacing, poses, and original audio. In <Video 1>, people appear as flat gray silhouettes. Use the silhouettes only for position, pose, and motion; never render gray shapes. Do not copy any people from it.
The rider silhouette in <Video 1> becomes <Subject 1>. The police officer silhouette in <Video 1> becomes <Subject 2>.

summary:
A playful two-person chase on the empty dusk road from <Image 3>. <Subject 1> rides in a crouched pose on an invisible moped while <Subject 2> follows on a futuristic hover-bike. 16:9, about 15 seconds.

retention_analysis:
Keep <Subject 1>'s face, hair, skin tone, build, and clothing from <Image 1>.
Keep <Subject 2>'s face, hair, and clothing from <Image 2>.
Use <Image 3> for the road, vanishing point, lighting, and time of day only. Do not take people, vehicles, or text from it.
Ignore the background, road, and vehicles of <Video 1>; take the setting only from <Image 3>.
Follow <Video 1> for the shot order, shot durations, camera movement, pacing, and music timing.

detailed_description:
<Subject 1> stays in a low seated crouch while riding: knees bent, feet lifted, and both fists held forward as if gripping invisible handlebars. No vehicle under <Subject 1> at any time, only empty space. No moped, motorcycle, scooter, or seat beneath them. <Subject 1> briefly leaves this pose for one clean mid-air flip, then lands and returns to the same crouch.

<Subject 2> follows behind on a futuristic hover-bike with two round red lights. Keep both subjects visually consistent and in the same roles throughout.

Shot 1 — Wide shot: <Subject 1> glides toward the camera along the road, grinning and leaning into the turns. The hover-bike follows behind.
Shot 2 — Close-up: <Subject 2> grips the hover-bike controls, focused on the chase.
Shot 3 — Medium shot: <Subject 1> performs one clean mid-air flip, lands safely, and resumes the crouched riding pose.
Shot 4 — Wide shot: both race away toward the vanishing point, with <Subject 2> still chasing.

Match each shot to the corresponding shot and duration in <Video 1>. Keep the same road, dusk lighting, and direction of travel in every shot. Match the action and cuts to the rhythm of the original video.

overall_soundscape:
Use the original audio from <Video 1>, including its music and synchronized sound effects. Preserve the original timing and rhythm. Do not replace it or add new music.

negative_prompt:
gray silhouettes, gray figures, moped, motorcycle, scooter, seat under the rider, extra people, police car, text, watermark, face swap between subjects

On this site the maker fills in your chosen length (15 seconds) and shape (9:16 or 16:9) on the summary line. Cartoon version adds one extra line asking for 3D animated characters that still match the two photos.

What it costs

No free render that covers a full clip, and no subscription. 15 seconds at 768P is 50 credits. Packs are on the pricing page.

FAQ

Do I need to type this in the maker? No. Upload two people, pick a background, generate. The prompt stays in the background.

What is Image 3? The empty road — a preset or your own still, with nobody in it.

Why Video 1? Shot order and pacing. Not the faces, not the place. You do not upload the original TikTok.

Can I copy this into another generator? Yes, with the same three stills mapped the same way.

Why doesn't rewriting the prompt fix the face? Change the photo.

Two photos, one empty road, four shots. Start a generation