Storyboard a Five-Second AI Clip Like a Shot, Not a Scene

Five seconds is not a scene. It is one shot — and the clips that work treat it that way. The failure mode I see most often in first attempts is a prompt that tries to cram a three-act structure into the length of a single camera move: a character enters, discovers something, reacts, and the camera cuts twice. The model obliges, and the result is a blur of half-completed beats.

A better mental model comes from storyboarding. Directors do not storyboard scenes; they storyboard shots, and a shot has at most three beats. Here is how to apply that to a 5-second AI clip, using the official samples on our Explore page as reference material — all four are real MiniMax H3 output from our own GPU, with the exact prompt and seed attached.

Beat-by-beat: the 1-3-1 budget

Divide five seconds into a 1-3-1 budget:

• Second 0–1: establish. The frame tells the viewer where they are. Nothing important moves yet, or everything moves gently. This is the beat people skip, and skipping it is why clips feel like they start in the middle of something.
• Seconds 1–4: develop. One motion carries the middle — the camera move, the subject's action, or both, but one dominant idea. Three seconds is enough for exactly one gesture to read clearly.
• Seconds 4–5: land. The motion resolves. The turn completes, the push-in arrives, the figure exits frame. A clip that ends mid-gesture feels truncated; a clip that lands feels intentional at any length.

The product turntable sample is the cleanest example of this budget in the wild: a ceramic bottle "completes a slow smooth turn" — establish the bottle, spend the middle on the rotation, land back near the starting angle. One subject, one motion, one light source ("soft key light"), locked camera. The prompt is two sentences long, and the clip works because the storyboard was decided before the prompt was written, not inside it.

Choosing the first frame

If you are doing image-to-video, the first frame is your establishing beat made literal — and it does more work than any words you write after it. Three rules from what survives contact with the model:

1. The first frame should already be a good still. If the composition is not right as a photograph, motion will not rescue it. Generate or pick the still first, judge it as a still, then animate.
2. Give the motion somewhere to go. A frame that is already perfectly balanced leaves the model nothing to develop. Slight asymmetry — a subject off-center, an empty area of frame — is an invitation for the camera or the subject to move into it.
3. Then say what must not change. The I2VA format exists precisely for this: "Preserve the mountain lake, dawn palette, shoreline, and framing… Add no people, buildings, or weather changes." The preservation clause is doing as much work as the motion description. Without it, the first frame is a suggestion.

The prompt builder on the MiniMax H3 prompt guide assembles this preamble for you in the browser, including the exact reference sentence the model expects at 0.00 seconds.

One continuous move beats three cramped cuts

Look at the four official samples and count the edits. There are none. Each is explicitly "one continuous shot":

• the aerial push over the coastal road at dusk — one drone push-in;
• the macro ink dispersing in water — one locked frame, the motion entirely in the fluid;
• the product turntable — one rotation, locked camera;
• the rainy night market walk — one handheld forward walk.

That is not a stylistic preference. It is what five seconds can hold. A cut at second two means the viewer spends a second re-establishing the new frame — a fifth of your clip, gone. When the H3 guide shows a two-shot prompt with a timed cut at 00:06.000, it is a 10-second example. The budget scales: ten seconds can hold one cut; five cannot.

This also tells you what to do when a clip comes back feeling empty: the answer is usually not a second beat, but a stronger single one. "Camera pushes in gently" is weak. "Camera tracks beside her at a steady pace" is a move with a speed and a relationship to the subject.

Studying the samples as storyboards

The Explore samples double as a study set because every clip ships with its full prompt, settings (5s · 768p HD), and seed. The method:

1. Watch first, prompt second. Play a clip with sound and write down, in your own words, when the motion starts and when it lands. Then open the prompt and find the clauses that produced those beats. The night market clip's "steady rain, footsteps, and distant indistinct market ambience" is why the sound feels like a place and not a soundtrack.
2. Map each clip to the five parts — subject, action, camera, light, style — from the prompt structure guide. Every sample maps cleanly. That is the point of publishing the prompts: the structure is checkable, not a claim.
3. Remix the settings, not the concept. Take the macro ink clip and swap cobalt for a different pigment, or the turntable for your own product shape. You are changing one variable against a known-good storyboard, which is the fastest way to learn what a clause actually controls — at 8 to 12 credits a draft on the local model, five experiments cost less than a coffee.

A five-second clip planned this way — one establishing frame, one developed move, one landing — reads as deliberate. The ones that feel like accidents are almost always three shots' worth of intent squeezed into one shot's worth of time. Storyboard the shot you can actually fit.