Video generation
Video is the most expensive place to be careless. Every mistake from image generation applies, plus motion, plus sound, plus the fact that a viewer watches it in real time and cannot skim.
11 ways to get it wrong
“make a cinematic video about our company values”
- 3.1
Write the prompt as if it were a still image.
WhyVideo needs a described movement. What moves, how fast, and does the camera move or the subject.
ResultA photo with a slow zoom on it. Everyone can tell.
- 3.2
Try to tell a whole story in one generation.
WhyCurrent models hold a few seconds of coherence. A story is built by cutting shots together.
ResultA clip where the character's jacket changes colour halfway through the sentence.
- 3.3
Ignore continuity between shots.
WhySame character, same wardrobe, same light direction. Without a locked reference, each generation invents a new world.
ResultA sequence that feels like a dream, when you were selling accounting software.
- 3.4
Never specify the camera.
WhyHandheld, dolly, static tripod, drone. The camera decides whether the shot feels tense, calm or cheap.
ResultDefault floaty drift on every clip, the visual signature of unedited AI video.
- 3.5
Leave sound until the end, or skip it.
WhyAudiences forgive rough visuals and abandon rough audio. Sound carries most of the perceived quality.
ResultA beautiful clip that feels cheap the moment it plays.
- 3.6
Generate at full length before testing the look.
WhyTest cheap and short, then commit. Compute spent on the wrong direction is gone.
ResultYour monthly credit budget, used up on take one of eleven.
- 3.7
Skip the storyboard because AI is fast.
WhySpeed of generation does not remove the need for a plan. It makes the plan more valuable, because you can execute more variations.
ResultForty clips and no edit. Nothing to assemble.
- 3.8
Put dialogue lip-sync on a face you generated from scratch.
WhySync errors of even a few frames read as uncanny to every human viewer.
ResultThe audience watches the mouth instead of hearing the message.
- 3.9
Deliver 16:9 to a vertical feed.
WhyFraming is platform-specific. A wide shot loses its subject inside a phone crop.
ResultTwo percent completion rate and a conclusion that 'video doesn't work for us'.
- 3.10
Never label synthetic footage of real events or people.
WhyTrust is the product. Once the audience finds out on their own, they reread everything you ever published.
ResultReputational damage that no content calendar repairs.
- 3.11
Use motion for its own sake.
WhyA pointless camera move tells the viewer that nobody made a decision here.
ResultMotion sickness and a skipped ad.
Grading your failure
Three levels. Anyone can reach bronze on a Monday morning.
The clip is pretty and communicates nothing.
Viewers drop at second three, every time.
A comment thread forms about the hands.