How to write a scene the AI actually follows
Who, where, what happens, what they say and how it's shot. A simple way to write scenes that turn into the clip you had in your head — with before and after examples.
The scene box is the whole camera crew. Whatever you write there is all the AI knows about your shot, so the difference between a clip that lands and one that wanders is almost always in the writing. The good news: it's a skill, and it's quick to learn.
The five parts
A scene that works usually answers five questions, roughly in this order:
- Who is in it — and what they look like.
- Where it happens — the place, the time of day, the light.
- What happens — one clear action.
- What they say — the exact line, in quotes.
- How it's shot — close-up, wide, handheld, slow push-in.
You don't need headings or a template. Just write it as a short paragraph in that order.
One beat per clip
The most common mistake is putting a whole story into one clip. A clip is a shot, not an episode.
- 5 seconds holds one beat: a look, a reaction, a single line.
- 10 seconds holds a setup and a payoff: something happens, someone reacts.
- 15 seconds holds a small scene: a bit of movement and a line or two.
If your scene has the word "then" in it more than once, it probably wants to be two clips.
Dialogue: short, in quotes
Put spoken lines in quotation marks and say who says them. Keep them short: a five-second clip comfortably fits about a dozen words. Long speeches get rushed or cut.
The wizard looks straight into the camera and says, flatly: "Have you tried turning the dragon off and on again?"
Say how it's shot
Naming the shot does a surprising amount of work:
- Close-up for reactions and lines.
- Wide shot for slapstick and anything with more than two characters.
- Handheld or phone footage for chaos and "caught on camera" energy.
- Slow push-in for drama — which is funny when the subject is a sandwich.
One rule matters here: on trololo.tv nothing may pass for real footage of a real event. A news-anchor parody or a "security camera" gag is fine when it's obviously a joke; making it look like something real happened is not. The content rules have the details.
Be concrete
The AI fills every gap with something average. If you want something specific, say it:
- not "a man in an office" but "a tired man in a wrinkled blue shirt at a desk covered in coffee cups, fluorescent light flickering"
- not "a funny dog" but "a golden retriever wearing tiny reading glasses, very serious"
Colours, props, materials and moods are all fair game. Describe what you want to see rather than what you don't — "an empty street" works better than "no people".
Use your pictures
If you attached pictures as references, they're numbered: Image 1, Image 2 and so on. Refer to them by name in the scene:
Image 1 sits at the reception desk of the hotel in Image 2, ringing the bell impatiently.
That's how you keep the same character from clip to clip — more on that in character sheets.
Before and after
Before:
A cat is a lawyer and it's funny.
After:
A grey tabby cat in a tiny black suit stands at a courtroom podium, wood panelling behind it. It knocks a glass of water off the podium, looks at the judge, and says calmly: "No further questions." Close-up, then the camera pulls back to show the stunned courtroom.
The second one is longer, but every word is doing a job.
A last tip
When a clip comes out wrong, don't rewrite everything. Change one thing — the shot, the action, the line — and generate again. You'll learn what each part of your writing actually controls, and you'll get there faster.
