AI fruit video prompt checklist: lessons from our banana-peel test

A useful AI fruit video prompt tells the model who is in the scene, what happens, and what the viewer should hear. Before adding more detail, check that those instructions fit together. A beautiful frame will not rescue a missing action or a reply assigned to the wrong character.

We can show what that looks like with a real example. On 10 October 2026, we recorded the same banana-peel comedy task across five models. This article uses the original Seedance 2.0 Mini take to explain how to review a prompt and its output. The sample is one take, not a model success-rate estimate.

If you are starting from an idea and want the complete production workflow, read How to make AI fruit videos. For more story ideas, use the copyable fruit story prompts.

Watch the example and read its original prompt

The scene has one joke. A banana tries to impress a strawberry, slips on a peel, and pretends it meant to fall. The strawberry does not believe it.

The recorded settings were Seedance 2.0 Mini, a requested five-second duration, 9:16, and 480p, with one reference image. These describe this recorded take; check the current tool controls for the options available when you generate.

Use the uploaded reference image for both characters and the kitchen. Preserve the banana's yellow tracksuit, the strawberry's red overalls, and both characters' fruit-head designs throughout.

Create a five-second comedy using the following four shots in order, with three distinct camera cuts. Each cut continues the same event.

Shot 1, approximately 0–1 second:
A medium two-shot of the banana and strawberry. The banana proudly flexes like a bodybuilder, looking at the strawberry with a smug grin. The strawberry watches skeptically. The banana is completely unaware of the loose banana peel on the floor.

Shot 2, approximately 1–2 seconds:
Cut to a wide, low-angle shot showing the banana's full body and the peel. Still looking toward the strawberry without noticing the peel, the banana confidently steps forward onto it. The peel slides under its shoe, causing its feet to slip forward. The banana loses balance and lands on its bottom.

Shot 3, approximately 2–3.5 seconds:
Cut to a medium close-up of the banana, now seated on the floor. It immediately gives a thumbs-up, forces a proud grin, and says in English, "Totally planned."

Shot 4, approximately 3.5–5 seconds:
Cut to a close-up of the strawberry looking down toward the seated banana. The strawberry chuckles, gives a sarcastic smile, and replies in English, "Sure it was."

Use quick slapstick timing and expressive reactions. Synchronize each character's mouth with its own spoken line.

Display clear, readable English subtitles at the bottom center of the video. Show "Totally planned." only during the banana's line, then "Sure it was." only during the strawberry's reply. Match the spoken words exactly.

Include a cartoon slipping sound, a soft landing thud, and the strawberry's chuckle. Keep both spoken lines clearly audible. Maintain consistent character appearances, clothing, kitchen layout, and spatial continuity across all four shots.

The original story prompt, recorded style instructions, and input reference are available in the linked comparison. The story text alone is not a complete record of every provider input. Our archived composed prompt was reconstructed from the app's composer rather than captured from the production request.

1. Make the action visible

"A banana is embarrassed" names an emotion. "The banana slips, sits on the floor, then gives an awkward thumbs-up" gives the model actions a viewer can recognize.

In this example, the loose peel needs to appear where the banana steps. The fall needs enough framing to show its body and the floor. A close-up of a surprised face alone would omit the event that makes the joke work.

Before generating, ask whether a viewer could understand the scene without reading your prompt. Write the cause and the reaction as visible actions. A kitchen is enough setting for this joke; extra locations and background events would add competing instructions.

2. Give each character a fixed identity

The reference shows a banana wearing a yellow tracksuit and a strawberry wearing red overalls. The original prompt asks the model to preserve those designs and the kitchen throughout.

Choose a few details you can keep unchanged across takes: the fruit, the outfit, and a recognizable face or expression. If you want a whole fruit head, describe that explicitly. A person wearing a fruit hat is a different design.

Use the image input mode supported by your selected model. A reference image guides appearance; a first-frame input specifies the opening image. The recorded Seedance take used a reference. Other takes in the comparison used different supported image-input modes, so the study is not a controlled comparison of identical input modes.

When you review the result, look at clothing, fruit shape, and who remains on which side of the scene. Reusing a reference helps communicate a design, but the output still needs that review.

3. Put the events in an order

The original prompt divides the action into four shots: showing off, slipping, an embarrassed recovery, and the strawberry's reply. Each shot continues the same event.

For a simpler first attempt, you can reduce that to one wide action shot and a reaction. The following is an adapted exercise, not the exact input that produced the published sample:

Create a short vertical 3D fruit comedy in one bright kitchen.
A banana in a yellow tracksuit shows off in front of a strawberry
in red overalls. Both have whole fruit heads with expressive faces.

Show the banana stepping on a loose peel, slipping, and landing
on its bottom. Keep its feet and the peel visible during the fall.
Then show the banana giving an awkward thumbs-up while the
strawberry watches skeptically.

Keep their clothing, fruit heads, and the kitchen consistent.
Use a playful slipping sound and a soft landing thud.

If your output skips the fall, simplify the scene before adding more adjectives. You might remove a cut, shorten the reaction, or give the action more time. Change the part that failed rather than expanding every part of the prompt.

4. Assign dialogue to the right speaker

The original scene gives each character one short line and names the language:

• Banana, after falling: "Totally planned."
• Strawberry, in response: "Sure it was."

That is more specific than asking for funny conversation. The prompt also requests that each mouth move with its own line and that both lines remain audible.

Check that the model you choose supports generated audio. In the result, listen for the words and the speaker, then check their timing against the visible action. Clear words, convincing acting, and accurate mouth timing are separate things to judge.

Native generated speech also differs from uploading a recorded voice track. Use a dedicated audio or lip-sync workflow when that is the task you need; a dialogue instruction does not establish that capability.

5. Check subtitles separately

The original prompt requests two exact subtitle lines at the bottom center, each shown during the corresponding speech. The published take has those lines, but you still need to inspect the result you receive.

Read the spelling and check that the words appear while the right character speaks. Also view the clip at phone size. A technically readable line in a large player may be too small when the video is shared.

If exact captions are required, correct them in an editor after downloading. Prompting for subtitles does not replace checking the actual lettering and timing.

6. Fix one observed problem

Our review found the intended setup, readable slip, and embarrassed recovery in this take. It also found an extra peel-like shape after the fall. That is a concrete continuity issue to notice even when the joke works.

Write down the defect before editing your prompt. For example, "The fall is visible, but the prop changes afterward" is more useful than "The video feels wrong." You can then choose an experiment: reduce the number of cuts, keep the floor in view, or simplify the prop instruction.

Those changes are experiments, not verified fixes for this sample. Keep the earlier result so you can compare the next take. If the new version removes the action while improving the props, you have found a trade-off rather than a finished clip.

The five-model comparison shows why output review matters. One shared story can produce different movement, staging, dialogue, and instruction-following across individual takes.

A prompt checklist to use before your next take

1. Can I name the main visible action in one sentence?
2. Can the chosen duration fit the action and the reaction?
3. Have I kept character descriptions and image references consistent?
4. Does each spoken line identify its speaker and language?
5. Do my framing instructions let the viewer see the cause of the event?
6. Have I checked the output's sound, subtitles, continuity, and ending?

When the clip works, download the MP4 and play the saved file. For a longer story, plan the next scene with the same character descriptions and references. The complete creation guide covers the workflow, and Fruit Movie Maker is the place to try your next scene.