Making it with AI · 11 min read
AI filmmaking: from script to shot
AI changes what a shot costs, not what makes it good. How to plan a microdrama so you spend where it shows: the workflow, writing a prompt a model can follow, keeping faces the same, and planning around what AI gets wrong.
AI video has made a shot cheap to try and easy to get wrong. It does not do the thinking for you. A model draws what you describe, so the quality of the picture depends on how clearly you know what you want: which size, which angle, which move, which light. The first four lessons are exactly that. This one shows how to use them when a model is your camera crew.
The workflow: decide cheap, render once
The mistake that costs the most is rendering expensive video before you know the shot works. Make the decisions in cheaper steps, and spend on the final render only when the earlier steps say yes.
- Script: the episode on the page, one beat per scene (lesson 1).
- Board: quick sketches of every shot. Cheap, so try several.
- Stage: a finished still of each shot with the real faces, places and light. Check the look here.
- Shoot: video, made from the staged frame. This is the expensive step, so you only shoot frames you approve.
- Voice: a read for every line, cast to each character.
- Edit: cut the shots together, add sound and captions, then publish.
Write a prompt a model can follow
A good shot prompt reads like a shot list line, not a story. It names the size, the angle, who is in the frame and what they do, the move, the light and the mood, and it says how long the shot is. Build one below: pick a choice from each group and watch the prompt assemble.
medium close-up. from a low angle looking up. Iona, nineteen, long braid and a plain linen dress, stops in front of the iron furnace gate and lifts her chin. slow push-in toward the subject. flickering warm firelight at dusk, amber highlights, cool stone shadows. tense, a held breath. vertical 9:16 frame, one continuous shot of about 8 seconds, no cuts.
How it is built
One line a model can follow
- Size and angle
- Medium close and low angle: where the camera is and how high, so the model frames the face the way you meant.
- Who does what
- One person, one action. Describe the look the same way every time so the character stays the same.
- Move
- Slow push-in: builds tension or realisation. One move per shot, so the model can hold it.
- Light and mood
- Firelight at dusk and tense: the feeling you want the audience to have.
- Length
- Say how long the shot is and that it is a single take. Most models hold best for a few seconds.
Shot size
Angle
Move
Light
Mood
Which model has what power
No single model is best at everything. Each one is strong at a few things and weak at others, which is why a studio picks a model for each shot rather than one for the whole episode. Press a model to see its powers, what it is for and what to watch for.
- Speaks its own linesYes
- Holds a character’s face—
- Follows camera moves—
- Shot length8 s
Video model
Veo 3.1 Quality
Final shots that speak their own dialogue.
- Used for
- The shoot step of the workflow.
- Watch for
- The most expensive of the video models, so use it for the shots that carry the episode.
Video Shoot: a staged frame becomes a moving shot
Images Board and stage: sketches and finished frames
Voice A read for every line, in every language
| Model | Speaks its own lines | Holds a face | Camera moves | Shot length |
|---|---|---|---|---|
| Veo 3.1 Quality | Yes | — | — | 8 s |
| Veo 3.1 Fast | Yes | — | — | 8 s |
| Sora 2 | Yes | — | — | 4 to 12 s |
| Kling 3.0 | — | Yes | Yes | up to 15 s |
| Seedance Pro | — | — | Yes | 5 or 10 s |
| Gen-4 Turbo | — | — | Yes | 5 or 10 s |
| Higgsfield DoP | — | — | Yes | 5 s |
| Hailuo 02 | — | — | — | 6 or 10 s |
| Wan 2.2 | — | — | — | 5 s |
Keep faces and places the same
A microdrama is a series, so the heroine has to look like the same person in every shot of every episode. A model does not remember, so you have to remind it each time.
- Character sheet: one approved picture of each person, from the front, side and back, used as the reference for every shot.
- Same words: describe a character the same way every time, such as the braid, the linen dress and the soot on her hands.
- Locked looks: choose a colour and light style for the series and reuse it, so scenes cut together.
- Place references: the same picture of the furnace gate for every shot set there.
What AI gets wrong, and how to plan for it
Every model has habits that show up again and again. Knowing them is how you plan around them, rather than paying to find out.
- Things that move on their own: a gate that should stay shut swings open, or a crowd walks the wrong way. Say what must not move, and cut away before it goes wrong.
- Hands and small objects: fingers, rings and phones drift. Frame them as a short insert, not a long close-up.
- Text on screen: signs, screens and captions come out garbled. Add real text in the edit.
- Long takes: most models hold best for a few seconds, so write shots of five to ten seconds and cut.
- Faces turning away: identity slips when someone turns. Prefer shots where the face stays toward the camera, or cut when they turn.
- Speech: mouths often do not match the line. Add the voice separately and sync the lips as a final step.
Voice and sound come separately
Most video models make a picture first and the sound afterwards, or not at all. In a microdrama that is an advantage: you can cast a voice for every character, record each line to its own timing, and replace a line without re-rendering a shot. Work the voices as their own pass, after the picture is settled.
Quick quiz · AI filmmaking
Question 1 of 4