When I first started writing prompts for AI video, I described what I wanted the final image to look like.
Something like:
A cinematic futuristic city at night, beautiful neon lighting, dramatic atmosphere, highly detailed, realistic.
It sounds descriptive, but there's a problem: almost everything in that prompt describes appearance.
Very little explains what should actually happen during the shot.
Lately I've been using a different approach. Before describing style, I write the camera and action.
Start With the Shot
For a simple product scene, I might begin with:
Medium shot. The camera slowly pushes toward a bottle standing on a table. The bottle remains centered while soft reflections move across the surface.
Only after that do I add the environment:
Dark studio background with soft directional lighting and subtle reflections.
Then I add style if it's actually necessary:
Clean commercial product photography, restrained contrast.
The complete prompt becomes:
Medium shot.
The camera slowly pushes toward a bottle standing
on a table. The bottle remains centered while soft
reflections move across the surface.
Dark studio background with soft directional lighting
and subtle reflections.
Clean commercial product photography with restrained
contrast.
This structure is much easier for me to edit than a long paragraph of adjectives.
The Four Parts I Use
Most of my video prompts now contain four sections:
SHOT
ACTION
ENVIRONMENT
STYLE
For example:
SHOT:
Wide establishing shot. Slow forward camera movement.
ACTION:
A small electric car travels along an empty road.
ENVIRONMENT:
Mountain road shortly after sunrise. Light mist in
the valley.
STYLE:
Natural lighting, realistic materials, quiet travel-film
look.
The advantage isn't that this is some magic syntax.
It's that each section has a job.
If the motion is wrong, I look at ACTION.
If the framing is wrong, I change SHOT.
If the scene feels wrong, I adjust ENVIRONMENT.
I don't have to rewrite everything.
Describe One Camera Move Clearly
Another habit I've dropped is asking for several camera movements in one short shot.
This:
The camera pushes forward, circles the product,
tilts upward and then quickly pulls back.
may describe the sequence I have in my head, but it's also asking the model to solve several motion instructions at once.
For an initial test, I'd rather use:
The camera makes a slow clockwise orbit around
the product while keeping it centered.
Once that works, I can decide whether the shot actually needs anything else.
Simple prompts are also easier to compare because there are fewer instructions competing for attention.
Write Observable Actions
I've also become more careful with abstract words.
Consider:
The character confidently enters the room.
What does "confidently" mean visually?
I can make the action more observable:
The character walks into the room at a steady pace,
keeps their head raised and stops beside the desk.
The second version describes things that could actually appear in frames.
The same principle works for objects.
Instead of:
The car moves dramatically.
I would write:
The car accelerates from a stop and passes the camera
from left to right.
Separate Shots Instead of Writing a Mini-Movie
Long prompts sometimes contain an entire edit:
Start outside the café, move through the door,
show the barista making coffee, cut to a close-up
of the cup and finish with the customer drinking it.
That's several shots disguised as one prompt.
I now split something like this before generation.
Shot 1
Wide exterior shot of a small café.
Slow push toward the entrance.
Morning light.
Shot 2
Medium shot behind the counter.
A barista pours steamed milk into a cup.
Static camera.
Shot 3
Close-up of the finished coffee on the counter.
Very slow push-in.
Soft window light.
Even if I eventually want one continuous sequence, thinking in shots forces me to decide what each part of the video is supposed to accomplish.
Keep the Model Replaceable
I've been applying this structure while testing different AI video workflows rather than designing prompts around one generator.
For example, LTX 2.5 is one of the video models I've been looking at while thinking about more controllable and open video workflows.
But I don't want my prompt library to depend entirely on one model.
My notes therefore look like this:
shot:
medium close-up
camera:
slow push-in
subject_action:
object remains stationary
environment:
dark studio
lighting:
large soft source from camera left
style:
minimal commercial footage
From there I can adapt the wording or parameters for whatever model I'm testing.
That also makes comparisons more useful.
A Template You Can Reuse
Here's the basic template I've settled on:
SHOT:
[shot size and composition]
CAMERA:
[one clear camera movement]
SUBJECT:
[what is visible]
ACTION:
[observable movement or event]
ENVIRONMENT:
[location, background, time]
LIGHTING:
[light direction and character]
STYLE:
[optional visual treatment]
Example:
SHOT:
Medium-wide shot of a desk viewed slightly from above.
CAMERA:
Slow lateral movement from left to right.
SUBJECT:
A laptop, notebook and ceramic coffee cup.
ACTION:
Steam rises gently from the cup. Everything else
remains stationary.
ENVIRONMENT:
Quiet home office in the early morning.
LIGHTING:
Soft window light from the left.
STYLE:
Natural, understated documentary look.
It isn't necessarily shorter than my old prompts.
It's just easier to reason about.
The Prompt Is Really a Shot Specification
The biggest change for me was stopping the question:
What words will make this video look impressive?
and replacing it with:
What should the viewer actually see during these seconds?
That leads naturally to framing, camera movement, subject action, environment and lighting.
Style still matters.
But now I add it after I've decided what the shot is supposed to do.
For video generation, that order has been much more useful than starting with a pile of visual adjectives.
Top comments (0)