How to make a short film with AI: plan it shot by shot

What making a short film with AI actually involves

This is the first lesson of Make a short film with AI, a free six-lesson course. By the end you will have a plan for a one-minute film that you can generate, voice and cut in the following lessons.

The most useful thing to understand first is the unit of work. Video models do not make films. They make shots: a single continuous clip, typically 4 to 10 seconds, sometimes up to 15 on newer models. A short film is dozens of those shots, generated one at a time and assembled in an edit, exactly like live-action footage. People who type a whole story into one prompt get one confused clip. People who plan shots get a film.

The clip above is a single shot from the example film we plan in this lesson. It works because it asks for very little: one character, one action (riding), one camera move (tracking alongside) and a setting that explains itself. The bear stays in the same place in frame for the full five seconds while the coast streams past, which is what a tracking shot should do. Every shot in your film should be this easy to describe.

The workflow this course follows has six stages, one per lesson:

1. Plan: logline, beat sheet, shot list, format and budget (this lesson).
2. Lock the look: style frames and a consistent character.
3. Animate: image-to-video prompts with deliberate camera direction.
4. Choose models per shot.
5. Sound: voice, music and effects.
6. Edit, pace and export.

From idea to logline and beat sheet

Start with one sentence. A logline names a character, what they want and what stands in the way or what changes. If you cannot write it, you do not have a film yet, only a mood. Ours:

Then break the minute into beats. For a 60-second short, five beats is plenty: each one gets roughly 10 to 15 seconds, which is two or three shots.

Notice what the story avoids. There is no dialogue, only two characters, and three locations. That is not a limitation of imagination; it is a choice that plays to the medium.

How to write a shot list for AI video

The shot list is where an AI film is won or lost. Each row should be something a video model can do in one generation. Four rules keep it that way:

• One action per shot. "He zips his jacket" is a shot. "He zips his jacket, grabs his keys and walks to the bike" is three.
• One camera move per shot. A push-in, a pan or a locked-off frame. Combining moves is where clips start to warp.
• 4 to 10 seconds. Most models generate 5 or 8 seconds well. Plan shorter shots than you think: edits are what make a film feel fast.
• Write the end state. Note where the shot finishes ("bike exits frame left", "he looks up"). It gives you a clean cut point and a clear motion prompt later.

Here is the full shot list for our one-minute bear film. Shot 5 is the clip at the top of this page.

Twelve shots, 61 seconds. Every row is a single verb and a single camera instruction, which makes each one a realistic generation. Shot 10 borrows the fox from the variable-matrix renders in lesson 2, where one prompt template turns a bear and a fox into frames that look like they belong to the same film.

Choose the aspect ratio and format up front

Decide where the film will live before you generate anything. Composition is baked into every frame, and cropping a 16:9 shot down to 9:16 later throws away two thirds of the picture and usually the subject's surroundings with it.

Check the ratios your chosen models support before you commit. Many video models offer only 16:9, 9:16 and sometimes 1:1. In image-to-video, the start frame's shape usually decides the clip's shape, so generate your storyboard stills at the final ratio too. If you need both a horizontal and a vertical version, plan them as two compositions rather than one crop.

AI storyboard from script: stills first, motion second

Rendering every shot straight to video is the most expensive way to find out your idea does not work. A still costs a fraction of a clip, so storyboard first: turn each row of the shot list into an image prompt, generate a few options, and only animate the ones you approve.

A grizzly bear in a leather jacket riding a vintage café racer motorcycle through the Pacific Coast Highway, cinematic golden hour, 35mm film photograph, slight motion blur, joyful

A grizzly bear in a worn leather jacket and aviator goggles rides a vintage café racer motorcycle along a coastal highway at golden hour, camera tracking alongside, fur rippling in the wind, waves crashing below, cinematic

These two renders are the same idea, a grizzly in a leather jacket on a café racer on the coast road. The still ran on Flux 2 Pro for 8 credits; the clip ran on Kling O3 Pro for 141. Look closely and you will also see they are not the same bear. The clip's prompt added goggles, but even apart from those the coat color and the face differ. They were generated from text on two different models with no shared reference, so each model invented its own bear. That is the consistency problem lesson 2 solves: it builds a character sheet for this bear and carries his jacket, collar and goggles into the garage and lighthouse shots of this film. It is also the main reason to animate from approved stills rather than generating every shot from text.

1. 01

   Write one image prompt per shot

   Describe the frame as it looks at the start of the shot: subject, framing, setting, light. Keep the style words identical in every prompt so the board looks like one film.

2. 02

   Generate three or four options on a cheap model

   You are testing composition and story, not final quality. An inexpensive model such as Flux 2 Pro (8 credits per image in our renders) is enough at this stage.

3. 03

   Lay the frames out in order and read the board

   Put the approved stills in sequence with their durations. Missing beats, repeated angles and confusing jumps are obvious here and cost nothing to fix.

4. 04

   Re-render finals with references

   Once the board works, regenerate the keeper frames at full quality with your character and style references attached, so every shot shows the same character.

5. 05

   Animate only the approved frames

   Each final still becomes the start frame of an image-to-video generation. The picture is already decided, so the motion prompt only has to describe what moves.

How to make AI video longer than 8 seconds

People ask this constantly about AI video, and the honest answer is that you rarely need to. Films are made of cuts. The average shot in a modern film is a few seconds long, and a one-minute short with ten to fifteen shots will feel better paced than one long generated take. When you genuinely need a longer continuous shot, you have four options:

• Pick a model with longer clips. At the time of writing, Kling 3.0 and Seedance 2.0 both generate up to 15 seconds, while Veo 3.1 tops out at 8. Longer clips cost more and give the model more time to drift, so check each model's limits on the catalog.
• Extend from the last frame. Export the final frame of a clip and use it as the start frame of the next generation, with a prompt that continues the motion. Some models build this in: Veo 3.1's scene extension generates each new clip from the final second of the previous one.
• Use first and last frames. Generate the start and end stills, then ask a model that supports both to animate the move between them. You control exactly where the shot lands, which makes the next cut clean. Lesson 3 shows one: the bear riding toward the lighthouse, landing on a frame where he has pulled up below it.
• Hide the join with a cut. Cut to a close-up or a reaction shot, then back. The audience accepts a jump in time between two shots far more readily than a visible seam inside one.

What an AI short film costs in credits

Budget a film the way a producer does: number of shots, times takes per shot, times price per take. The price per take is the variable that moves most. To see how much, we rendered a version of shot 4 (the bear riding out of the garage) from the same start frame, with the same motion prompt, on five video models:

That is 57 to 305 credits for the same five or six seconds of footage: Seedance 2.0 Fast cost roughly 5 times as much as PixVerse V6. Applied to our twelve-shot film, with three takes per shot, the arithmetic looks like this at the prices we paid:

Most films land somewhere in the middle rows, because not every shot needs the premium model. Wide establishing shots, landscapes and simple motion often look fine on an inexpensive model. Save the expensive renders for the shots carrying the story: faces, hands, hero moments. Lesson 4, choosing the best AI video model, covers which model to use for which kind of shot.

In PromptLab, the Playground shows a live credit estimate before every run and failed runs are not charged, so you can keep a running total against your shot list as you go.

Shot list and prompt template you can copy

Keep one card per shot. It doubles as your prompt sheet: the image prompt makes the start frame, the motion prompt animates it, and the notes keep continuity across the film.

Fill in the shot list, write a card for each row, and storyboard it. Before you animate anything, you need the bear in shot 1 to be the bear in shot 12. That is the subject of the next lesson: how to keep characters consistent in AI video.

Questions people ask