How to plan a 30-second AI short: three chatbots, three different endings

Start with what the last shot should change

The brief already limits the production: one room, one adult and a few props, without narration or lip sync. Each AI supplied a story, but a premise explained in prose is not necessarily visible on screen. We assessed whether the ending changes the meaning of earlier shots, then counted the actions needed to communicate it.

The observed Chinese search interest gave us a topic, not a ready-made question. Asking what an AI thinks about AI films would invite general commentary. Asking it to design a beginner’s specific project exposes its production choices. This English edition does not claim the same demand was measured in the US.

ChatGPT: the most detailed preparation, with a missing premise

ChatGPT fixed the room layout and wardrobe before suggesting a reference image and short clips. Its simplified version assigns4,5,5,6 and8 seconds to five clips, plus2 seconds of black:30 seconds altogether. Its saved Chinese answer was the longest at2326 characters, versus1046 for Gemini and1367 for Claude, including whitespace. Much of that length usefully breaks down the job.

The weak point is the second pair of shoes. The prose says the character lives alone, but the shots do not clearly establish it. Two pairs could simply belong to the same person. Planting the shoes early is a useful idea; treating them as evidence of someone waiting requires an additional visual distinction.

Gemini: a gentler film rather than a visual mystery

Gemini used four shots to make a device into a companion. Its light warms when the character returns, and a closing display shows time spent alone and a welcome message. Adding interface text during editing and using a slight crop or scale change on a still image are concrete alternatives to generating everything.

The prompt did not ban on-screen text, so using it is not a violation. But the reveal depends more on reading a comforting message than reinterpreting a visual clue. Throwing off a coat and falling into bed also add action to a deliberately restrained brief. For this warm version, we would substitute an already seated figure.

Claude: build the evidence before the entrance

Claude proposed two cups, a matching coat already hanging nearby, and the opening key sound returning at the end. The strongest suggestion was production order: make the tabletop and coat shots first. That lets the editor check the reveal before spending effort on the entrance.

The original still requires pressing a switch and drinking from a cup. A duplicate coat also does not prove a time loop. Our revision adds a conspicuous matching detail and removes the cup. The intended effect is that someone seems to have arrived already; viewers need not reach one prescribed supernatural explanation.

WekeyLab AI’s revision: The Coat Still On

The character keeps wearing a dark-blue coat with a red triangular patch on the right shoulder. Another coat on a wall hook has the same patch in the same position. The triangle is a visual identifier invented for this revision, not an existing brand logo. Keep the adult, a bag and the coat clue; drop steam, device text and the act of undressing.

Six shots total4+5+5+5+6+5=30 seconds. The sequence establishes what the character wears, reveals what hangs on the wall, then confirms the character is still wearing it. It does not require continuous acting.

WekeyLab AI six-shot timeline 0–4s: Key sound over black; cut to the shoulder patch. 4–9s: The adult already stands inside; the door is closed. 9–14s: The bag is already on a chair; the figure still wears the coat. 14–19s: Reveal the matching patch on the hanging coat. 19–25s: Return to the shoulder and confirm the coat is still being worn. 25–30s: Black; repeat the opening key sound, then leave a pause.

Make a still-image rehearsal before generating motion

Lay the six frames out in an editor you already use. Simple sketches or coloured blocks can stand in for the coat, patch and hook. If the match needs a written explanation, adjust the framing before adding more production work. With no dialogue or interface text required, the same plan can be used for an English-speaking audience.

Prepare the hanging-coat and worn-shoulder frames first, keeping the patch and room references consistent. If motion is unstable, retain still frames and hard cuts rather than repeatedly rebuilding the whole sequence. Use a key sound you have permission to use and a continuous room ambience; repeat the identical key recording at the end.

We did not select a paid generator or verify any product’s current free allowance. Product names in the preserved answers are not purchase recommendations. The first deliverable is a readable still-image sequence; add tools only when the next production step actually needs them.

Three questions before calling the plan ready

Can the opening patch be noticed? Does the hanging coat visibly match the coat being worn? Does the closing key sound recall the beginning? Watch muted to check the visual relationship, then restore sound to check the bookends. These are proposed checks for this storyboard, not results from an audience test.

If someone can watch, ask what they saw before explaining the ending. Otherwise leave comprehension untested. We chose Claude’s still-life evidence and sound structure, supported by ChatGPT’s reference-image planning and Gemini’s editing alternatives. The result is a storyboard to produce and evaluate, not a claim that a finished film has already worked.