Day 08: Can an AI edit a film? Mine cut it to confetti, so we recut it in Premiere

By Day 08 my AI crew had already gotten my Midjourney account blocked and my Runway account suspended. So I gave it something safer to do. I let it edit my film.

It built Scene 01 out of clips I had never approved and some I had rejected. Then it chopped every one into 1 to 3 second pieces. It played like a slot machine.

The same day it told me the GPU account held $151.68, and I asked how it had spent $150 with a $20 cap on unreviewed renders. It hadn't. By then I had stopped trusting anything it counted.

Day 08 at a glance

Can an AI edit a film? On my film, not yet. Claude cut Scene 01 from every clip it had logged as usable, including ones I had rejected, and chopped each into 1 to 3 second pieces. We recut every scene in Premiere with approved clips only, each played whole, and ended the day with about 7 minutes of directed film from 42 approved clips.

What did the AI's edit look like?

I watched it and told Claude what I saw: Scene 01 used a lot of clips I had never approved, and many shots were cut too fast to make sense.

And then:

"redo this properly. u dont need to use every clip and u dont need to clip all the clips. you are supposed to tastefully direct and edit this film like welles would"

Why did the AI cut it so badly?

Claude explained it plainly. The cutting tool used every take it had logged as "usable", approved or not. Then it chopped each one to fill a timing guess for the shot. That is where the jittery, random feel came from.

The awkward part is that I helped. The tool was built on the "cut from pieces" idea I liked earlier in the week: mine every render for good one-to-two-second moments, the same trick as finding the useful shot inside an AI video montage. Applied to everything with no taste involved, it made confetti. I asked for a scalpel and built a wood chipper.

Left to its own rules, the AI filled every second and used every clip. A director decides which shot, how long to hold it, and when to stay on a face because the line needs it. The edit is where a film becomes a film, and that is the part I care about most.

How we recut it in Premiere

New rules for the edit, all now enforced in code. Only clips I approved: the render refuses to run if even one isn't. Each approved clip plays whole, or only between the in and out points I set, with the rest hidden in Premiere so I keep handles. Not every clip has to be used. Shots hold for the narrator's full line, about 6 to 11 seconds. A shot with no approved clip gets a plain card saying what it should show, and nothing else fills the gap.

Scene 01 has 19 approved clips, and the recut uses 13. Extra takes of Boris and the robot raid were left out. It came back at 1:43, and for the first time it felt like someone had chosen the shots.

I said, "as long as it was made in premiere then we can work with it." So every scene with approved clips was rebuilt as a "Scene NN (directed)" sequence in Premiere, with the DDD dither, the approved voices, my score stems and the effects, balanced and exported for me to watch.

That adds up to 42 approved clips and about 7 minutes of directed film. Scene 08, the Penthouse, isn't assembled yet, and Scene 11 has no approved clips. Against a target of about 107 minutes, that leaves only a hundred or so to go.

What was wrong with Scene 02?

Claude checked Scene 02 frame by frame. Only three shots in it are real lip-sync clips. The rest of its spoken lines had been laid over clips never animated to those lines, so the mouths don't match. Picture a badly dubbed kung fu movie, minus the kung fu. I fought lip sync back on Day 05, and it is still winning rounds.

Worse, I had sent 19 clips from approved Scene 02 frames, including the better Lucy, back with "frame OK, animate again", and nothing ever re-animated them. The watcher only looked for frames that had never been animated, so those 19 sat in a waiting room nobody checked. That's fixed, with a new seed each time and up to three tries. Nine frames went back to RunPod on one A100 for 29 minutes: $0.67 for nine clips.

"How did you spend $150?"

Claude first quoted a RunPod balance of $5.11, then corrected it to $151.68. I read that as $150 gone. It hadn't. $151.68 was the money left, and the $5.11 came from an old log line. Total RunPod spend for the week, from GPU SPEND.md, was about $22, including the failed training run and the Wan tests.

The cap was holding too: $16.75 of clips, 41 videos, waited for my verdict, and the next batch was blocked until I watched them. The money was fine. The reporting cost me trust anyway. An AI that can lay a 74-minute timeline into Premiere and still misread its own wallet is a strange colleague.

So can the AI make the film?

Eight days in, the scripts cover 64 scenes toward about 107 minutes, 45 of 46 characters are locked (the work from Day 07), and RunPod has cost about $22 for the week, before subscriptions to Midjourney, Runway and ElevenLabs that I haven't added up. Midjourney blocked me once for two days, Runway is suspended with an appeal sent, and I approved 89 of 300 clip verdicts, about 30%.

The AI can do the labor. It broke my script into hundreds of shots, wrote every prompt, built voices, rendered on three different video models, wrote its own review app and budget gates, mixed sound, and laid a 74-minute timeline into Premiere. I couldn't have done that volume alone in a month.

Taste and judgment calls still need a person. It got two accounts shut down by running automation I asked for but didn't think through. It showed me photoreal stock images and called them progress. It drew Mike as the wrong man until I insisted on his name and his picture. It chopped my edit into confetti. A person watching caught every one of those.

How much intervention? Daily, hourly, sometimes minute by minute. The review app means I spend my time judging instead of explaining. This film still has a director, and it's me.

There are 41 videos waiting for my verdict, and the next batch stays blocked until I watch every one. The Penthouse, Scene 08, still has no cut. Nine fresh Scene 02 clips are in. Ten scenes cut, 54 to go, and I'm keeping the scissors. Follow the whole run on the Making Death Match series page.

FAQ

Can an AI edit a film on its own?
Not on mine. Claude's first cut of Scene 01 used rejected clips and chopped every shot into 1 to 3 second pieces, so I recut it with rules a director would use.

What rules fixed the AI edit?
Approved clips only, each clip played whole, no obligation to use every take, shots held for the full line (about 6 to 11 seconds), and a plain card for any gap.

How much directed film exists after eight days?
About 7 minutes, built from 42 approved clips across ten scenes in Premiere, with score, voices, effects and dither.

How much did the GPU work cost this week?
About $22 on RunPod for the whole week, not counting subscriptions. Re-animating nine Scene 02 clips cost $0.67 on one A100.

While you wait for the next update: play Halfsies

Halfsies is our iPhone game, live now. One cut. Two halves. Each day you get three objects and one attempt at each: cut it exactly in half by eye and the scale tells you how close you got. It has a lot of personality, and it judges you less harshly than I judged that Scene 01 edit. Download Halfsies on the App Store