The workflow the best AI filmmakers use

Six stages, in order

Tools change every few months. The order of work barely moves. Filmmakers who publish their process keep landing on the same sequence, and it looks a lot like traditional production: write, break down, design, shoot, cut.

1. 1. Story

   A premise, a visible choice, and beats that fit your runtime.

   You leave with: A beat sheet you can read aloud in the time the film runs.

2. 2. Sequences

   Split the beats into clips. Each one is a single video generation.

   You leave with: A list of sequences with a duration and a job for each.

3. 3. Look

   Describe each character, place, and prop once. Make reference images.

   You leave with: One fixed sentence and one approved image per subject.

4. 4. Start frames

   Compose the first frame of each sequence with an image model.

   You leave with: A still per sequence that already looks like your film.

5. 5. Video

   Animate each start frame with a staged prompt. Generate several takes.

   You leave with: Takes you have watched for story first, then continuity.

6. 6. Edit

   Cut the chosen takes. Add sound, music, color, and titles.

   You leave with: A finished film, and notes on what you would change.

Decide where it is cheap

Text costs almost nothing to revise. A still image costs a little. A video take costs the most, in credits and in waiting. So the workflow pushes every decision to the cheapest stage that can hold it.

Settle the ending while it is a sentence. Settle the character’s face while it is an image. By the time you generate video, the only open question should be motion and performance. If you are rewriting the story inside a video prompt, go back a stage.

1. Start with a story you can see

A premise is ready when a character does something visible and that action changes something. Work with an LLM as a writing partner, but judge the options yourself. Our guide to developing a film idea takes you from a blank page to a story worth making.

2. Break it into sequences

Plan in units the model can actually generate. One sequence is one generation, usually 4 to 15 seconds, with a clear start and end state so its neighbors line up. A sequence can hold more than one shot when the prompt stages them.

Give every sequence one job. If you cannot say what the audience learns in it, merge it or cut it. Narrative to sequences to prompts covers this layer in depth.

3. Lock the look before you generate

A video model has no memory between clips. Consistency has to come from you. Write one sentence per character, place, and prop that names only what the camera can see. Reuse it word for word in every prompt that shows the subject.

Words keep a description stable. They do not keep a face stable. For that you need reference images: generate them once, approve them, and attach them wherever your tool accepts references.

4. Make start frames first

Image-to-video gives you far more control than text-to-video. Compose the first frame of each sequence with an image model, using your reference images. Fix framing, light, wardrobe, and set here, where a retry takes seconds.

Lay the start frames out in order and read them like a storyboard. If the film does not make sense as stills, motion will not save it.

5. Generate video, then review like a director

Now write the video prompt: what moves, how the camera moves, and what we hear, split into timed stages. Our cinematic prompts guide has the vocabulary and a full one-take example.

Expect to generate more than one take. Watch each once for story and once for continuity. Change one instruction at a time, starting with the biggest failure, and keep the prompt saved beside the take it produced.

6. Finish in the edit

Generated clips are raw footage. Trim the soft first and last half second. Cut on action. Replace or layer generated audio with sound you chose on purpose. A film with average shots and good sound reads as finished. The reverse does not.

One film through all six stages

Here is the 30-second cleaner film from our idea guide, carried through the workflow. It is an illustration, not a record of a production.

Story

A night cleaner about to erase a chalk hopscotch game tries one hop and leaves it.

Sequences

Five clips of 4 to 8 seconds. The hop gets its own close clip.

Look

“The cleaner: a tall figure in grey coveralls pushing a yellow cart.” Reused word for word. One image of the cleaner, one of the school corridor.

Start frames

Five stills. The chalk grid is drawn once and matched in every frame.

Video

Three takes of the hop. Keep the one with a convincing landing.

Edit

Cart wheels, a scuff of chalk, no music until the last shot.

What goes wrong when you skip a stage

• No story: beautiful clips that do not add up to anything.
• No sequences: prompts that ask one generation to do a whole scene.
• No locked look: a lead character who changes face every clip.
• No start frames: credits spent discovering the composition in video.
• No review pass: continuity errors you only notice in the edit.
• No sound design: a film that still feels like a demo.