My wife and I star in a one-minute film set on the Dhaka metro. We have never once been on the Dhaka metro.
No camera. No crew. No shoot day. Two prompts, thirty seconds each, and the film was rendered.
I have run a story-driven animation and video studio since December 2017. EndingScene has delivered 1,000+ video projects for 500+ clients. I know what a shoot day costs in people, permits and patience. So when I say the render was the cheap part, I am not being modest about the tools. I am telling you where the work actually was.
The confession: my first AI video attempts were terrible. Beautiful, and terrible. Dragons, drone shots, a face that changed between shots. I was doing what everyone does in week one: opening the generator first and hoping. It took me an embarrassing number of credits to accept that the tool was not the problem. I had skipped the script.
The render is the cheap part
A commercial shoot needs a director, a camera operator, a lighting person, an art director, models, an editor. Eight to ten people at least. In AI video that unit still exists. It now lives inside your prompt. If the prompt has no director in it, you get footage with nobody in charge.
That is why my rule is story first, render second. Not one frame before the story is locked.
Story first: the seventeen-shot script
For the metro film, Claude wrote the story first: seventeen shots with timecodes. Not "a couple on a train". Shot one, zero to four seconds, wide, platform, morning light, she looks at the departures board. Shot two, close, his hand on the pole. Each shot with camera angle, movement, lens feeling, lighting, duration.
This is a shot-list, the oldest tool in filmmaking, and it is the thing that makes AI video work. A generator cannot decide what matters in a scene. A script can.
Character sheets and locks
Two more steps sit between script and render, and both exist to stop the model from improvising.
Character sheets. ChatGPT built them from our own photographs: face, hair, clothing, the bag she carries, written as a fixed description that every shot reuses. The consistency problem in AI video is not "can it make one good shot". It is "can it keep the same person, the same product, the same mood across five shots". A reference image and a written sheet are the only answer I have found.
Locks. The prompt carried locks: this outfit, this light, this camera height, no new characters, no new locations. A lock is a sentence the model is not allowed to negotiate. Without locks the second shot drifts, and by the fifth shot you have a different film.
Then the render: two prompts, thirty seconds each. Boring, and that is the point.
The system, in order
- Write the story as a shot-list with timecodes. Four to six shots for a reel, seventeen for a minute.
- Build character and product sheets from real reference images. One image per person or product, reused in every shot.
- Write the locks: what cannot change between shots.
- Generate. Two shots first, check them side by side, then the rest.
- Edit on one table: footage, voice, music, captions.
- Check every frame before anyone else sees it. Hands, fingers, text on signs. AI still gets these wrong.
One line I will not cross: a real person's face or voice only with their written permission. Past that line it is not content, it is deception.
The tool will change again next quarter
Everything I just described is tool-agnostic on purpose. The generators I used this month will not be the ones I use in six months. The shot-list, the sheets, the locks, the frame check: those have not changed since the first commercial I made. AI moved the price of the render. It did not move the price of thinking.
In my studio, AI-assisted production has cut delivery timelines by roughly 40 percent. Not because we generate faster. Because the script was always the slow part, and now the script is where all our time goes.
Practice
What to do on Monday
- Pick one product or service and write a fifteen-second, four-shot script with timecodes. Do not open a generator.
- Choose one reference image for the main subject and write a one-paragraph character or product sheet from it.
- Write three locks: what stays constant in every shot.
- Generate only the first two shots and put them side by side. If the subject drifted, fix the sheet, not the prompt.
What is the last video you made where the script took longer than the edit?