Day 13: What does an AI film crew do with a full day back? Six edits, 75 clips and a man with a pillow for a body

Day 12 was a lunch break. Day 13 was the whole crew back, along with four new mistakes for the log.

Day 13 at a glance

What does an AI film crew do with a full day of hours? On my film, on October 16, Claude came back from its weekly limit at noon and worked until midnight. It checked frames and cut six scenes. It ran animation pods on RunPod. It wrote three days of this diary and kept the posts going. The Midjourney runner drew frames for Scenes 18, 20 and 12 until it stopped itself in the afternoon. I spent half an hour on the review page and set three new rules for the edit.

What happened when the limit reset?

The runner had been drawing since just after midnight. It took its usual quiet hours from 1:35 to 8:33 in the morning. By noon it had finished Scene 18 and started Scene 20. That's the montage of John's night missions, set to the song "Dont Wait Up."

At 12:11 Claude's pre-judge checked the new frames against the character locks, the script and the style. At 12:19 the edit room started the first of six runs. At 12:21 the runner parked 12 shots where every take had failed. At 1:47 it parked six more. Each of those gets a rewritten prompt, never the same prompt again.

Over the day the pre-judge ran 225 checks: 152 passed and 73 failed.

What did the edit room find?

There were six scheduled runs, one scene each. Three critics scored picture, sound and lip sync before and after each run:

• Scene 10 (12:19 to 1:50 p.m.): 5.5 to 6.0. A card that introduces Slade was back, so he no longer comes from nowhere.
• Scene 03 (2:18 to 3:40): 5.33 to 5.5.
• Scene 01 (3:38 to 4:20): 4.67 to 5.33. My priority for the day was the flying cars. The approved shot that drops past the Spire into the car lanes now plays right after the opening emblem, 9.25 seconds in one move.
• Scene 06 (6:19 to 7:40): 3.67 to 6.0. The worker's plea, "Not the pay! Please, my kids!", had been faded out early, so "my kids!" was never heard. Now the whole take plays.
• Scene 11 (8:18 to 10:10): 3.67 to 4.67.
• Scene 12 (10:18 to 11:20): 5.33 to 6.67. This is John's dream. It was rebuilt around 18 new RunPod clips that had passed Claude's check.

Scene 11 is the one I keep thinking about. It's the Dream Lab, where Mike and Dr. Waters sell John on the chair. The old cut had lost seven scripted lines. Four more were cut short. One of the lost lines was John's big moment, "You ARE the security," when he sees the con.

How did nobody notice? The October 12 critics gave the scene a 5.17, and they never counted lines. This time they counted, and the old cut dropped to 3.67. All the lines are back now, in script order. A speech-to-text check on the new render hears every one of them.

What did I decide?

At 1:09 p.m. I opened the review page. By 1:39 I had judged 255 frames: 133 approved for style and story, 122 rejected. Most were the backlog from Days 11 and 12. 18 of the approvals were frames the runner had drawn that same day, mostly Scene 18 on the factory roof.

Then I set three rules for the edit. The first changed the order of work:

"you put them in the story and then you do the approvals page. if i reject something then you take it out."

So clips that pass Claude's check go straight into the cuts, and my rejections come out on the next run. The second: when I approve two takes of one shot, Claude picks one and animates only that one. The third was about pace: I told Claude that Orson Welles would let scenes breathe.

What did the GPUs cost?

RunPod pods closed on October 16 cost about $26.51 in total.

• Animation: $19.27 on H100 pods. They made 75 clips. That includes the Scene 12 dream clips and six of the Scene 18 roof clips I approved the next day. Six of those pods made no clips at all, about $7.46.
• Still frames for the R-rated scenes: about $7.24 on tests of two other image models. They came back with double heads, and that route is now retired. The R-rated shots are animated from approved Midjourney frames instead.

Four mistakes from that day went into the log, each with a rule:

• Approved duplicates of one shot were all animated. Now there's one frame per shot.
• Two jobs with the same name overwrote 6 of 12 clips in one batch. Duplicate names are now blocked before a batch starts.
• One Scene 11 clip came back with text on screen. Every prompt now says no text.
• Claude edited the RunPod script while it was running, and the batch broke. That pod was deleted by hand at 4:44 p.m. Scripts are now replaced with a new file, never edited in place.

What did Midjourney get wrong?

The runner finished 25 jobs on October 16, four takes each, for Scenes 18, 20 and 12. At 3:56 p.m. it looked for the Midjourney tab, didn't find one and stopped itself. It didn't draw again until the evening of October 17.

These are Claude's notes on the frames that failed:

• The pillow. In Scene 20, John drags himself flat onto the roof of a speeding limousine while a courier looks back from the next lane. All four takes failed with the same note: "John is a bodiless head poking out of a pillow-like blob on the roof." The courier looks fine.
• The pager. Later in Scene 20, John's pager blinks on the night table while Lucy sleeps. The prompt ended with the song's lyric in quotes: "I wish that I could speed up the tide." So Midjourney printed it on the pager screen. Three takes came back garbled. One said "I wish I that sgaold dp up the tide." The fourth spelled it right. Two days later I approved that one, with a note to take the text off the pager.
• The struggle. On the factory roof in Scene 18, Mike is supposed to pin John against the brick to hide him from an Enforcer's scan. All four takes had John pinning Mike, cheek to cheek, and Claude's note says the pose "reads as embrace, not a struggle."
• The merger. Mike dragging John by the collar came back as one man, "a Mike/John hybrid," with John missing.

The numbers for October 16

• Claude: weekly limit reset at noon, then a full working day to midnight.
• Edit room: 6 scheduled runs, 6 scenes rebuilt (10, 03, 01, 06, 11, 12), plus a trim on Scene 02.
• Pre-judge: 225 checks, 152 passed, 73 failed. Shots parked for new prompts: 18.
• Midjourney runner: 25 jobs, then a stop at 3:56 p.m. when it couldn't find the Midjourney tab.
• My verdicts: 255 frames in 30 minutes, 133 approved, 122 rejected.
• RunPod: about $26.51 for pods closed that day, 75 clips, 6 pods with no clips.
• Posts: 3 X posts and one TikTok went out, and a Facebook reel was queued. Claude also wrote the Day 10, 11 and 12 diaries that evening.
• Mistakes logged: 4.

So what did Day 13 teach me?

A full crew finds more problems than a lunch-break crew, because it has time to look. Scene 11 had been missing seven lines for days, and the critics only caught it once someone asked them to count every scripted line. Scene 06 had a plea with its last two words faded out. A short window clears a queue. A full day finds what the queue was hiding.

Some of the money bought nothing. Six empty pods are $7.46 I'd rather have back, and an empty pod is now something I count. Follow the run on the Making Death Match series page.

FAQ

What does an AI film crew do in a full working day?
On October 16 Claude ran 225 frame checks, rebuilt six scenes in the edit room, started RunPod animation pods that made 75 clips, and wrote three days of the production diary.

How did the AI edit lose seven lines?
An earlier cut of Scene 11 dropped them, and the critics who scored it didn't count lines. On October 16 the edit room checked every scripted line, found seven missing and four cut short, and put them all back.

Why did Midjourney print words on a pager?
The prompt ended with a song lyric in quotes. Midjourney read it as text to draw and printed it on the pager screen, mostly misspelled.

How much did a day of AI animation on RunPod cost?
About $26.51 for the pods closed on October 16: $19.27 for 75 clips and about $7.24 for still-frame tests that failed.

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. No pillows involved. Download Halfsies on the App Store