Short films made
with AI, made
properly.

A boy is walking his dog in the park and looking the wrong way. Benny catches something on the air, and the leash is out of the boy's hand before he registers it. What follows is a one-minute chase that ends the way the dog always intended: nose to nose with a golden cocker spaniel, and — accidentally — two children saying hello to each other for the first time.

Twenty shots, made for eight-to-thirteen-year-olds, at roughly a minute. The story is simple on purpose. The difficulty was never the idea — it was continuity. Four characters who have to stay recognisably themselves across every cut. One sun, locked on the right of frame, so that every shadow in the film falls the same way. A park that has to still be the same park sixty seconds later.

Shot 20 took nine attempts — not because the shot was hard, but because two dogs kept refusing to be the right size next to each other.

That one turned into the most useful thing I learned all year. I measured the height ratio off approved footage, stated it numerically in the prompt, and the render still read wrong. A correct height ratio doesn't guarantee a correct perceived size: coat volume and body bulk make a spaniel read bigger than the number says it is. Nothing in the wording fixed it. It got solved by hand, in an edit pass, after the generator had had its nine goes.

One shot was cut outright after six rejections and its beat folded into the shot after it. An entire environment got rebuilt three times — trees removed, then a path removed, then the camera raised to eye level — because a low reference plate was quietly making finished shots read as flat. Every job has a UUID. Every rejection has a written diagnosis. That log is the reason the film holds together, and it's the part of the process I'd bring to anyone else's project.