Two hundred and seventy characters, and the only prompt here whose settings fal published too.
This is the prompt fal puts in its own API example, which makes it the one entry in this library where the parameters are published alongside the text: 768P, 16:9, five seconds, balanced expansion. It is also the shortest complete lesson in camera grammar on the page — one move, named with its direction, and a lens that explains why the highlights behave the way they do.
Reference render — not generated on this site. Source: fal.ai
Output reference
1 FREE CLIP · NO AUDIO1 FREE CLIP · Agnes Video V2.0 · 16:9 · 5s · NO AUDIO
The clip beside the verbatim source. Copy it into the console above, change what you need, generate.
The camera move first, then the subject, then the lens and the light, then three sounds.
Slow dolly forward down a rain-slick night-market alley, steam pouring off a noodle cart under a buzzing red neon sign. 35mm anamorphic, practical lights only, wet cobblestones doubling every highlight. Sound: wok sizzle and clang, rain ticking on tin awnings, neon hum.
What is safe to change. Most libraries publish only this list, which is why so many copied prompts come back worse than the original.
A night market is a light source with people in it. Any replacement that keeps wet ground and practical lights works the same way — a fairground, a rain-soaked forecourt, a lantern-lit lane. Move it to daylight and every clause about reflections stops paying for itself.
"Slow dolly forward" names a type and a speed. Swap it for "slow truck left" or "static hold" and the shot changes completely while nothing else in the prompt has to. What you cannot do is remove the speed word and keep the type.
35mm anamorphic sets the frame’s width and the shape of its flares. Changing it to a long lens compresses the alley and shrinks the depth; changing it to a wide sets the cart further away. It is one of the cheapest words in the prompt and one of the most visible.
Wok sizzle and clang, rain on tin, neon hum. Three sources, all of them things in the frame. Adding a fourth is safe; replacing them with a mood — "atmospheric night sounds" — hands the mix back to the model.
Three ways to break it
"[Push in]" and "[Truck left]" are Hailuo 02 syntax. MiniMax H3 Max reads them as ordinary text rather than as instructions, so the brackets end up describing nothing and the camera picks its own move.
Five seconds and a continuous dolly leave no room for a cut. A second location forces one, and the model either compresses both into an unreadable pass or ignores the instruction.
That clause is what keeps the light motivated by things in the shot. Without it the render is free to add key lighting from nowhere, and the wet-cobblestone reflections stop matching anything visible.
Two hundred and seventy characters is short enough to read in one breath, and it still makes six decisions. That ratio is what makes it the most useful prompt on this page to copy the shape of.
It opens with the camera. That is deliberate and it is worth adopting. The official prompt writing guide puts camera near the end of the within-shot order, but a prompt this short has only one shot, and leading with the move tells the model what kind of clip this is before it has decided anything else. Everything after reads as detail attached to a dolly rather than as a still that something later has to move through.
The move itself is written the way movement has to be written for this model: as plain English with a type and a rate. "Slow dolly forward" has both. "Cinematic camera work" has neither, and a phrase with neither leaves the model to pick a type, an amplitude and a speed on your behalf — three decisions you did not make.
The lighting clause is the one that repays study. "Practical lights only, wet cobblestones doubling every highlight" is two constraints in eight words. The first bans light sources that are not in the frame. The second tells the model that the ground is a mirror, which turns every neon sign and every cart lamp into two light sources instead of one. That is why a five-second clip from this prompt looks lit rather than rendered.
The sound half is three items long and every one of them is a thing you can see or infer from the frame: the wok on the cart, the rain on the awnings, the sign that is already described as buzzing. Because H3 Max generates audio in the same pass, sound tied to visible sources is sound that lands in sync with them.
And because fal publishes the parameters with this one, it doubles as a settings reference. 768P is the resolution the model is tuned around. Five seconds is the floor. Balanced is the expansion mode that decides per request how much to rewrite you, which keeps the total time close to the render time — quality can spend up to thirty seconds on the rewrite before rendering starts.
Everything here runs in text to video, on MiniMax H3 Max.
It is fal’s API example rather than a showcase card, so it arrives as a code sample with resolution, ratio, duration and expansion mode filled in. The showcase entries publish the prompt text only.
For a prompt already written this precisely, yes. Balanced tidies and fills obvious gaps in about a second. Quality rewrites much more and can spend up to thirty seconds doing it, which on a five-second clip is longer than the render.
You can, and it is the right resolution for deciding whether the move is the one you want. 768P is what H3 Max is tuned around, so finish there.
No upload, no image pass. Copy the text and generate.
All prompt ideasCopy it, change three things, run it.
Every character is on this page. 5s at 768P costs $0.40.
Generate free · queueWritten and maintained by the MiniMax H3 Max AI Video Generator editorial teamPublished Last updated