One impressive shot doesn't tell you whether AI can do motion design. I wanted to know if it could direct, which means pacing a film, holding back and deciding when to cut. So I gave an AI agent a 15-second showreel (well, eventually I went up to ~30s-ish for more 'content' feel) to direct and kept every render. It took ten versions and just over two days.
How every version got checked
The whole film had to be one self-contained HTML file, which made rendering predictable. The same input gave the same frame every time, so I could render any single moment and have a program check it. Every version had to pass those checks before I'd watch it: cuts on the beat, lines on screen long enough to read, words brighter than everything around them, no dead beats (a beat where nothing on screen changes) or silent hits, and a clean loop. When I disliked something, I could ask whether a number had moved, and usually it had.
The first five versions
V1 passed every check and its cuts were on the beat, but I found it flat. It was too slow and too subtle, with plain backgrounds, no heat and some visuals that missed the beat entirely. Nothing in the checks said so.
V2 was denser and faster, and nobody could read it. The median shot was about a fifth of a second, so every word vanished on its own cut and three words shared less than a second. A 3D ring world took nine of the thirty-six shots, and stepped echo extrusions on most words made the type look fake. Since that round, 3D is for places and objects, and type stays flat and crisp.
V3 was the director's cut. It passed everything, and it's the one I'd throw out first. It followed V2's pacing rules, which forced forty cuts and five full-frame flashes. It also dropped V2's rule that an orange dot appears in every shot, so nothing carried the eye between cuts and every frame felt standalone.
The checks had fixed the last film's problem and created a new one. A green run only proves the film satisfies the checks, and I don't trust one for anything I still have opinions about.
V4 put the dot back and ran one continuous camera through the 3D journey. It came out smoother and less engaging than V3, whose beat-driven frames held attention better. It also rolled the picture about 175 degrees by 2.15 seconds, swung back, then rolled again around 2.75. A roll that size is a broken camera, and catching it is the clearest reason the checks exist.
V5 is the version I liked most, and I kept it as the short loop. It also broke my rule for public copy, which is no counts and no locations. The poster showed real counts and the HUD showed real map coordinates on every frame. To a viewer, a number like that reads as a claim about my work, and I can't defend it.
An outside review said the biggest beat dimmed the poster. It also called the landing echo a gray drop shadow, which misreads a three-frame arrival smear that only shows in a still. I kept the smear, and that's a taste call I own.
Turning a feeling into numbers
V6 ran thirty-nine seconds over the full chorus and had the strongest art direction so far. My note was that the momentum wasn't there, and parts of the film waited. The audit found fifteen dead beats, the worst a five-beat run from 29.143 to 31.286 seconds. That audit was the most useful thing in the project, because it turned "the pacing feels off" into a work order with timestamps.
V7 kept V6's art direction and rebuilt the four scenes where the audit and my own ranking agreed the film was weakest. The impact frame became a full-frame negative, used once in the film.
V8 moved seven moments onto the beat the music actually plays, instead of the beat the beatmap implied. I'd do that pass much earlier now.
V9 was nine small changes, written down first so nothing tempting crept in. The trail had been retuned twice and still rendered as a string of dots, so I had it deleted. Two others were plain bugs. The staircase built 61 treads for 27 entries and kept going after the ball had gone, an off-by-thirty-four bug. The rooftop word list had 44 entries against a count of 45, so one word showed up twice at once.
The burst through the glass, the biggest moment in the film, had no sound at all, and the music there is quiet enough to leave room for some.
The climb from about 13 to 15 seconds felt like it stalled when I wanted it to accelerate, and I couldn't explain why. The curve asked for a speed about 9 percent past the fastest it could ever reach, so the curve itself forced the stall. I replaced it with a speed profile built first and the position worked out from it, so the speed can never go backward.
V10 has 27 shots and 2 dead beats against V6's 15. Two cut checks still fail and the contrast check ends in FAIL. I left them, because in V6 making the accents big enough pushed the reading contrast into unreadable territory, and readability wins.
Where the AI was strong and where it wasn't
The AI was strong wherever there was a right answer, from beat alignment and contrast to finding a 175-degree roll or an off-by-thirty-four staircase in seconds. It was weak at notes with no right answer, like too slow, too subtle, chunky, hurts the eyes and the words look pasted on. Each of those notes made the film better, and the suite was already green every time I gave one.
A note on the videos
The ten files above shipped as 1080p60 masters at 14–22 Mbps — about 690 MB in total, enough to run this site's monthly hosting bandwidth out in days. They're now encoded as H.264 at CRF 23 with faststart: visually identical at viewing size, roughly 4.5× smaller. The original masters are archived offline.