We are two people and a gimbal. That is the entire unit, and an AI video maker is now the third member of it in every sense that matters except carrying things. Everything below came out of eighteen months of trying to make genre shorts look like they cost more than they did, and most of it is the result of getting something wrong first.
Week One: We Stopped Storyboarding on Paper
The old routine was pencil thumbnails taped to a wall, which worked fine right up to the moment somebody had to interpret them. Two people can look at the same box with an arrow in it and picture completely different camera moves. We found that out on a rooftop at two in the morning, which is a bad time to find anything out.
Now the thumbnails come out of an AI video maker before anyone leaves the flat. Not as final footage — as an argument-settler. If the shot reads on a phone screen at eleven at night, it will read on the day. If it does not, we have lost four minutes instead of a location fee.
This is the part people get wrong about bringing an AI video maker onto a production: the output is rarely the deliverable. It is the thing that stops two collaborators describing different films to each other while nodding.
The Night the Cape Changed Colour
Our first serious attempt at previz produced eleven versions of a hero landing, and the cape was a different red in nine of them.
That was our fault, not the model’s. We had written a paragraph of atmosphere and exactly one clause about the character. The lesson took a while to sink in, because it feels like the opposite of how prompting is supposed to work: the more evocative your writing, the less control you have. Every adjective competes. A model has no way to know that “crimson” is a continuity rule and “brooding” is a mood note unless you separate them.
So we split the brief in two. Locked column: character, wardrobe, props, location, anything that must survive across shots. Loose column: tone, weather, lens feel, the stuff we would happily let drift. It is not glamorous and it fixed about eighty percent of our problems.
The Order You Write Motion In Matters More Than the Words
Second thing that changed our hit rate, and this one surprised me.
Describe motion the way it happens in time. Opening state, then the main action, then what the camera does, then what changes in the scene, then the closing frame. Written in that order, a request behaves. Written as a pile of simultaneous impressions, it produces a lovely four seconds of something adjacent to what you asked for.
MiniMax H3 is the one we settled on for this, mostly because it reads text, image and audio references as a single set of instructions rather than three separate switches. Also released under the name Hailuo 03, which confused us for a week and may confuse you too. When we hand it a chronological sequence and a reference image doing one clearly stated job, the hit rate goes up enough that we stopped counting.
One AI video maker is much like another at this level. Veo, Kling, Wan and the rest are all in the same building now. What is not interchangeable is whether the person typing has thought about time.
Where the Method Broke
Full disclosure, because a piece like this is useless without it, and because the version of this article where everything worked would be a lie.
Previz stopped helping the moment we tried to use it for blocking with two actors. Interactions between people are exactly where generated video is least reliable and where a chalk mark on a floor is still undefeated. We wasted a weekend proving that.
It also does not survive contact with a location you have not scouted. A generated interior is an idealised interior. Ours had a window on the wrong wall, and we cheerfully built a shot around it before anyone visited the actual room. The gimbal does not care about your beautiful render.
And a small practical trap worth thirty seconds of your attention: on a free tier the clip normally comes back stamped, and the default visibility on your results is often open rather than private. For previz nobody cares. For a character design nobody outside your group chat has seen yet, check that setting before you upload the reference, not after.
One more that cost us a Saturday: we assumed a previz clip could double as a pitch asset. It cannot. A test made to settle a camera move looks exactly like what it is, and sending it to someone who has not been in the conversation reads as sloppy rather than efficient. Keep the two jobs separate, and rebuild the shot properly when it is going in front of anyone who matters.
Your Homework Before the Next Shoot
Take the single shot from your current project that you and your collaborator argue about most. The one where you keep saying “no, like, it sweeps” and they keep nodding without agreeing.
Write it twice. Once as you would normally — atmosphere, feeling, references to films you both love. Then once in strict time order, with a locked column and a loose column, and one reference image that has exactly one job.
Run both through the same AI video maker. Put the results side by side on a phone at arm’s length.
If the second one is not obviously closer to the film in your head, ignore everything above and go back to pencil. If it is, you have just replaced an argument with a four-minute test, and that is roughly the whole pitch.
