The AI filmmaking workflow, step by step

A workflow is only useful if each step produces something concrete that the next step uses. This guide breaks AI filmmaking into seven stages, and for each one describes the document or asset it should produce, what "done" looks like, and the mistake that most often sends people back to redo it. It is written for a short film, but the same chain scales to a music video, a commercial or a pitch trailer.

If you are new to the topic, start with the overview in AI filmmaking: a practical guide and come back here when you are ready to plan a real project.

Stage 1: logline

Write the whole film in one or two sentences before anything else. A good logline names a character, what they want, what gets in their way and what is at stake. "A lighthouse keeper who has not spoken to anyone in a year finds a message in a bottle addressed to her, written in her own handwriting." That sentence already tells you the main location, the main character, the key prop and the tone.

Done when you can read it to someone and they can picture the film. Common mistake: describing a vibe instead of a story ("a dreamy journey through memory"). Vibes produce montages; stories produce films.

Stage 2: script

Expand the logline into scenes. Use standard screenplay conventions if you know them: scene headings (INT. LIGHTHOUSE - NIGHT), action lines in present tense, and dialogue. For a first AI short, one to three pages is right.

• Write only what can be seen or heard. "She remembers her mother" cannot be generated. "She turns the bottle over; her mother’s ring is tangled in the cork" can.
• Keep the cast small. Every additional recurring character multiplies the continuity work.
• Favour locations you can reuse. Three scenes in one well-designed location will look better than three locations designed in a hurry.
• Decide early how dialogue will be handled: on-camera lines need lip sync, while voice-over and off-screen dialogue do not.

Done when every scene has a clear purpose and you could cut none of them without losing the story. Common mistake: writing action that needs dozens of shots to show, like a chase or a fight, into a film you intend to make in an afternoon.

Stage 3: shot list

The shot list is the bridge between the story and the models, and in AI filmmaking it matters more than in live action, because each line will become a prompt. Go through the script scene by scene and list every shot you need to tell it.

For each shot, record: a number (scene and shot, like 2C), the shot size, the camera angle, the camera movement if any, the subject, the single action, and roughly how long it should hold on screen. Add a column for notes such as "match to 2B eyeline" or "needs last frame".

Think about coverage as you go. Dialogue scenes usually need a wide or two-shot, then singles on each character, and the 180-degree rule still applies: keep each character on the same side of the frame across the scene. Plan a few insert shots of hands, props and details. They are easy to generate and they give the edit room to breathe.

Done when you could hand the list to a stranger and they would know what each shot shows. Common mistake: two actions in one shot ("she opens the bottle and reads the letter"). Split it. The model handles one action far better, and the cut between them is free storytelling.

Stage 4: look development

Before generating scene frames, decide what the film looks like and fix it in writing and in images. This is where you prevent the drift that makes AI films feel like a slideshow of unrelated clips.

• Mood board. Ten to twenty reference images that share a palette, a light quality and a texture. Use them to agree on the look, not to copy.
• Style line. One sentence you append to every prompt, such as "muted teal and amber palette, soft overcast light, 35mm film grain, shallow depth of field". Write it once and paste it exactly.
• Character references. For each recurring character, a set of approved images from several angles and a locked written description. See how to keep characters consistent.
• Location references. A wide establishing frame of each location, plus a couple of reverse angles, generated before any action happens there.

Done when you can generate a new frame in the film’s world and it looks like it belongs. Common mistake: changing the style line midway because one shot looked nicer with different words. Consistency beats local improvements.

Stage 5: storyboard

Generate one still per shot with an image model, using the shot list line, the character and location references, and the style line. Put the frames in order and read them as a sequence. Does the story make sense with the sound off? Do eyelines match across cuts? Does the light stay consistent within a scene?

Fix everything you can at this stage. Regenerating a still is quick and costs far less than regenerating a video. If a shot is not working as a still, it will not work as a clip either. How to storyboard a film with AI has the detail.

Done when the boards tell the story on their own and you would be happy if each one were the first frame of its shot. Common mistake: treating the storyboard as rough and planning to fix things in video. It almost never gets easier later.

Stage 6: video generation

Now animate. For each shot, give the video model its approved storyboard frame as the first frame, and write a prompt about motion: what moves, how fast, and what the camera does. The appearance is already set by the image.

• Choose the model per shot. One may hold faces better in close-up, another may handle a big camera move or a stylised look better. The models page lists what each one supports.
• Use a last frame when the shot must end somewhere precise, such as a character reaching a mark or a reveal landing on a composed frame. See first frame and last frame.
• Match cuts deliberately. To make one shot continue into the next, take the final frame of the approved take and use it as the first frame of the next shot.
• Generate a few takes and choose. Keep notes on which prompt and settings produced each keeper.

Done when every shot on the list has at least one usable take. Common mistake: rewriting the whole prompt after a bad take. Change one thing at a time so you learn what actually made the difference.

Stage 7: edit, sound and finish

Assemble the takes in shot-list order in your editor, then start cutting for rhythm rather than for completeness. Trim the soft first and last moments of each clip. Cut on action where you can: a turn of the head or a door swinging shut hides the join between two generations.

Build the sound in layers: dialogue first, then ambience, then effects on key moments, then music. If dialogue is on camera, generate the speech, then apply lip sync to the clip. Grade the whole film at the end so every shot shares the same colour world, add titles, and export at the resolution your platform expects.

Done when someone who has never seen the script can watch it and follow it. Common mistake: falling in love with a beautiful shot that slows the story. Cut it and keep it for the trailer.

Running the workflow on one canvas

The chain above works in any set of tools. The reason we built Pindow as a canvas is that stages four to six are much easier when the references, the storyboard frames and the generated clips sit on the same board. On Pindow you can place reference images next to an image generator and attach them as references, turn an approved frame into a video generation with that frame as the start, give supported models a start and an end frame, and extract the last frame of a clip to begin the next shot. The Prompt Engine’s presets help you write the style line and camera language consistently across a whole project.

When you are ready to put the workflow to use on a real project, how to make an AI short film covers choosing a story that suits the medium, estimating generations and finishing the film.