Scarlet Pictures
All journal entries

How to Plan an AI Short Film Before the First Generation

  • AI
  • Filmmaking

The most useful public brief for an AI short this month came from a jury.

At the first Astana AI Film Festival, Latvian filmmaker Nikolay Shestak won the $450,000 (about €401,000) Grand Prix for The First Honest Day. Euronews describes it as a film that “explores how people’s actions shape who they become,” with characters’ looks changing as their deeds do. The premise gives the picture something it can show. Shestak said, as quoted by Euronews:

Honestly, I didn’t expect to win. There were so many incredible works and different genres here that I’m curious myself why the jury chose my film.

In the main competition, second place went to India’s Abhik Banerjee for SITARA, and third to Spain’s Ernest Desumbila for Looking For Perseida. Euronews puts the field at more than 8,000 submissions from 125 countries, 25 finalists, 19 genres, and more than 80 generative models. Those films were judged on artistic vision, storytelling, and overall coherence, and on how organically the technology sat inside the idea. A plan is where that coherence can start.

On AIF 2026, Runway’s AI Festival asks for film submissions of 3 to 15 minutes, generative video, and a fully contained linear story. Where it prints a runtime, the winners run from Lucas Levitan and Fabián Jiménez’s Little Mes at 4:19 to Sifang Chen and Yang Wang’s Between Before and After at 8:21, with Dorian & Daniel’s The Well at 5:29 and Robert Gaudette’s A Face Only A Mother Could Love at 7:50 among them. A shorter story, finished and coherent, already competes.

Here is one way to plan an AI short film on a page before the first generation. The ferry story is a worksheet for the boxes.

Give the short a length you can finish

Pick the length before the look. Every extra minute is more shots that must share a face, a coat, a room, and a time of day. A 15-minute film can be beautiful work, and it is a lot of continuity on a first pass. About 5 to 8 minutes sits with the printed AIF runtimes, and it fits a want, an obstacle, and a turn.

Write the target in minutes, then as a shot ceiling. At about four seconds a shot, six minutes is on the order of 90 shots. Treat that as a ceiling. Drop a subplot while the cut is still free, and the same coat is more likely to survive to the last frame.

Write one want and one turn

A contained story needs someone who wants something, a reason they cannot have it yet, and one change that decides the ending. The AIF introductions show how small that engine can be. The Well follows a sharp-witted hare in a snow-covered trailer park who doesn’t buy the official story of a wolf found dead at the bottom of a well. One place, one question. YUUUKI’s Postman follows someone delivering a single old letter after the culture of letters has vanished.

Say yours in one sentence before any prompt. If it needs a second plot, you still have two films.

On the last ferry of the night, a deckhand finds a violin and has until morning rehearsal to get it back to the musician who left it.

Want: the musician has the violin before rehearsal. Obstacle: the last boat has gone, and the terminal is closing. Turn: she leaves the violin with the harbor office, plus a recording of the first line of the set list, so the owner can hear that it is safe. A radio, a case, and a voice can play the whole turn.

Lock the three things that have to match

Drift is usually a continuity break: the coat, the face, or the room changed. Write the marks a viewer uses as “must not change.”

Person. Three anchors. Call her June: cropped silver hair, a yellow oilskin coat, red knit gloves. Brown hair or blue gloves reads as someone else.

Prop. A black violin case, a white luggage tag, a strip of blue tape on the handle. The tag says “rehearsal 8am” and a first name. If the tape jumps to the lid, the audience feels it.

Place and time. One covered dock, one sodium lamp, rain already stopped, the ferry tied. Night stays night. Blue hour is a different scene, and it needs its own line.

The review of A Face Only A Mother Could Love shows character sheets, narration, and inserts carrying a finished short. The sheet here comes first.

Approve three stills before any moving clip: June front-facing in even light, the case on a plain ground, and the dock from the angle you will open and close on. If a clip drifts, regenerate the clip and leave the stills alone. Matching the reference to a stray frame is how one character becomes two.

Build the shot list with a continuity column

For each shot, write the image, the action, what must match, and what may change.

  1. Wide. June crosses the gangway left to right. Match hair, coat, gloves, lamp, and ferry. Change nothing. This angle returns at the end.
  2. Insert. Her red glove stops on a padlock. She has seen something in the cabin.
  3. Wide, inside. The case sits small on a bench. Match the blue tape, the white tag, and the sodium light.
  4. Close on the tag, “rehearsal 8am” and a first name. If the letters wobble, her voice reads them, so the fact survives.
  5. Medium. June on the dock with the case, facing the same way as shot 1, so the dock does not flip.
  6. Insert. A radio, her glove at the edge. We hear the call to the harbor office. The mouth can stay off screen.
  7. Close. The case opens on a penciled set list. If the pencil will not read, she says the first title.
  8. Medium. Phone on a cleat, violin under the lamp, one line of music. If a glove comes off, it stays off until you show it put back. If the fingers will not hold the note, cut to the case and let the audio play. A later pass can restore the hands.
  9. The reply, as a voice note over the dock. On-screen text can replace it once type renders cleanly. The voice note protects the ending now.
  10. Wide, shot 1’s angle. She locks the gate. Same coat, hair, lamp, ferry, and screen direction. The case is gone.

Ten shots are the spine, not the whole six minutes. Extra inserts need the same match column, so they keep the coat she already has.

Leave room for a stronger pass

Staying at the dock is a strength. The Well keeps its mystery to a trailer park and a well. The First Honest Day, in the Euronews description, makes one idea visible.

Two doors stay open. The reply begins as a voice note, and can become words on a screen when text is reliable. Shot 8 lets the recording carry the feeling, and June’s face can hold more of the reply later, when a performance will sit still. One idea that could make a plan like this land even harder: give the musician one vocal habit, and plant it when the case opens. When the reply arrives with that habit, the ending rhymes.

Copy this page before you generate

  • Length in minutes, and a shot ceiling.
  • One sentence: who wants what, and by when.
  • The obstacle and the turn.
  • Three anchors for the person.
  • One hero prop, with the marks that keep it the same object.
  • Place, time, and the light.
  • Three stills first: person, prop, master angle.
  • Numbered shots: image, action, must match, may change.
  • A backup wherever the beat depends on readable text or a precise hand.

Then generate. Order does more than the choice of model. The field Euronews describes used more than 80 of them, and coherence was still something a jury could reward.

The same question, what a shot has to hold, sits behind the generative experiments on the AI page.

Sources

The short films are on the films page.