For a long time, my response to a bad AI-generated video was simple: rewrite the prompt.
If the camera moved too fast, I changed the camera description. While I was there, I would usually tweak the lighting, add another sentence about the subject, remove a few words I didn't like, and generate again.
Sometimes the next result looked better.
Great — except I had no idea which change actually helped.
After repeating this enough times, I started treating prompts less like creative writing and more like something I could debug.
I started with five fields
Most of my test prompts now begin like this:
SUBJECT:
A pair of wireless earbuds on a white display stand.
ACTION:
The case slowly opens. The earbuds remain stationary.
CAMERA:
Slow push-in. Fixed camera height. No orbit.
LIGHTING:
Soft daylight from the left with a subtle shadow.
STYLE:
Clean commercial product shot with a neutral background.
There's nothing special about the wording.
The useful part is that each section has one job.
A few weeks ago, for example, I had a product shot where the camera kept moving farther around the object than I wanted.
The original instruction was:
CAMERA:
Cinematic camera movement around the product.
My first instinct was to rewrite the whole prompt.
Instead, I left everything else alone and changed only this:
CAMERA:
Slow push-in only.
Fixed camera height.
No orbit.
No handheld movement.
The next result wasn't perfect, but it answered a useful question: the camera instruction had been too open-ended.
That was much more useful than simply getting another random variation.
One prompt accidentally contained two experiments
Another mistake took me longer to notice.
I had written:
The camera slowly circles the shoe while the shoe rotates.
It sounded perfectly clear when I wrote it.
But when the result became unstable, I couldn't tell what I was debugging. Was the camera movement causing the problem? Was it the rotating object? Or was the combination too much?
So I split the test.
SUBJECT MOTION:
The shoe rotates slowly clockwise.
CAMERA:
Static medium shot.
No camera movement.
I tested the object rotation first.
Only after that looked reasonable did I add camera movement back.
It required an extra generation, but at least the second test told me something.
I've started doing the same thing with lighting and style.
I also use fewer words like "cinematic", "epic", "dynamic", and "professional" now.
Not because those words are always useless. They're just difficult to debug.
If I want a particular lighting setup, I'd rather write:
Strong side light, dark background, visible rim light.
than:
Dramatic cinematic lighting.
The first version gives me something concrete to change later.
The template beside my test folder
After a while, the same structure kept appearing in my prompts, so I saved a blank version:
SUBJECT:
[What needs to remain visually consistent?]
ACTION:
[What should the subject do?]
CAMERA:
[Camera position + one movement]
LIGHTING:
[Direction and quality of light]
ENVIRONMENT:
[Background and surrounding objects]
STYLE:
[Visual treatment]
AVOID:
[Problems observed in previous tests]
The last field used to be much longer.
I had copied a huge negative-prompt list from an older project and kept carrying it forward. Eventually I realized I couldn't explain why half of those terms were there.
Now AVOID starts empty.
If a test produces a recurring problem, I add something relevant. Otherwise I leave it alone.
My notes are intentionally boring
Beside each generation I keep a tiny text file.
A real entry from this kind of test looks roughly like this:
Test 07
Changed:
camera
Before:
slow orbit around product
After:
static camera + slow push-in
Result:
product shape more stable
less background distortion
Next:
keep camera
reduce subject rotation
It takes maybe thirty seconds to write.
That little file becomes surprisingly useful once there are ten or fifteen videos sitting in the same folder.
Before I started doing this, I had files called things like:
video_final.mp4
video_final2.mp4
video_final_good.mp4
I wish I were joking.
A week later, those filenames told me absolutely nothing.
I tried keeping the template when switching models
I was curious whether this approach was only useful with the generator I had been testing, so I started keeping the prompt structure unchanged when trying other video models.
One of those runs was with Wan 3.0.
I kept the same subject/action/camera structure rather than writing a special prompt for the model. That made differences easier to spot because I wasn't changing the model and the entire prompt at the same time.
Some wording still needed adjustment, which I expected. But I had a baseline instead of starting from an empty text box.
That's the part of the workflow I've found most useful.
When a generation looks wrong now, I don't immediately ask myself how to write a better prompt.
I open the last test note and look for the smallest thing I can change.
Top comments (0)