Day 01: How to make an animated film with AI, starting from my own screenplay

Before the week was out, my AI film crew had gotten me blocked by Midjourney and suspended by Runway.

On day one it was still on its best behavior. It only wanted to make an animated movie where nobody's mouth is ever on screen. It animated shots that floated like ghosts. Runway's filter rejected its opening shot twice. And it drew a hand with six fingers and called the shot done.

I'm the director. On day one that mostly meant saying no. By midnight I had said it a lot, and then I said the one yes I'd regret.

Day 01 at a glance

I'm making DEATH MATCH, an R-rated, 1980s-style animated feature, with Claude as my whole crew. On day one the AI wrote a 29-shot shooting script, broke 13 scenes into 228 shots, designed voices in ElevenLabs, built still frames in Midjourney and passed its first lip-sync test in Seedance 2.0. By 10pm it had cut 26 of 30 shots of one scene. I reviewed almost every frame for about 18 hours.

This is the first entry in Making Death Match, a daily diary of the whole production, failures included.

What is DEATH MATCH?

DEATH MATCH is a film I wrote. A dystopian city, a factory worker named John, his girlfriend Lucy, his friend Mike, a crime boss called Victor in a penthouse, and a tournament where people die on camera.

I want it grown-up and grimy, closer to the first Blade Runner than to anything for kids. The score is my own music. I have stems for eight songs.

Who does what: the human director and the AI crew

Claude is an AI coding agent that runs on my Mac. It writes and runs code, drives a web browser, calls APIs and edits files.

I'm the director. I wrote the script, made the visual seed of the style with my own Midjourney images, cast the voices, and I approve or reject every frame and every clip. Claude does the rest: shot lists, prompts, image and video generation, voices, sound, assembly and the edit.

So the diary asks one question. Can an AI turn a human's screenplay into a film, and how much human does that take? Fair warning: this is not a story where the robot wins. Two of my accounts got blocked before the week was out, and both times the AI's automation did it.

Where we started

I opened the session at about 6am, which is early for a director and normal for a robot. My desktop already had a folder called Deathmatch-handoff: a partial screenplay as a PDF, a few approved character designs from an earlier attempt, and my style references.

My first note was a complaint. The approved Elektra had visible AI artifacts on the left side of her sheet. Claude replaced her with a new design, S22b, which is now the approved Elektra.

Then I explained what went wrong last time. The earlier test felt like an animatic. A still frame would hold, the lip sync was bad, and every shot hard-cut to the next.

Claude's first plan was to keep mouths off camera so lip sync wouldn't matter. An animated film where nobody is ever seen talking. Bold. I told it: "Try not to avoid it, fix it." That sentence ran the rest of the week.

How do you keep one look across an AI film?

You write the look down once and reuse it on every frame. Ours comes from Midjourney v8.1 with my style reference and personal profile: smooth airbrushed cutout shapes, no outlines, a frosted glow, coral orange and teal, and a dappled film grain. Claude saved that as a prompt suffix.

I didn't want the look to drift from scene to scene. Keeping it from drifting turned out to be most of the job, and the AI's favorite way to fail.

Late in the day Claude built a finishing pass I call DDD, "dithered dotted diffusion". Try one looked like cheesecloth over the picture. Try two looked like a 1990s screensaver. I said keep going, and try three was right. My note at 5pm: "the dithering thing now looks great!"

What the AI did on day one

Here's the labor list. A small studio would split this across a dozen people. Claude did all of it before dinner, mistakes included:

• Shooting script. It rewrote the Penthouse section (Victor's office, about 3 minutes 45) as a 29-shot camera script with a floor plan, in a Welles style: low angles, deep focus, staging in depth.
• Scene breakdown. A second Claude session did the opening movement in parallel: 13 scene files, 228 shots, 27.9 minutes, zero validator errors.
• Scene kits. Empty-room plates of the penthouse from six directions, props, then a master frame for each camera setup, all sorted into contact sheets for me to pick from.
• Voices. Victor, Elektra and the Spy were designed in ElevenLabs. I picked one take of each, and chose John, Lucy and Mike from shortlists the same day.
• Cast. I needed a new John and a new Mike. Claude made option sheets and I picked John 1, Lucy 1 and Mike 3.

Does AI lip sync work for animation?

On day one, for one line, yes. Claude took a still of Victor and one of his recorded lines and ran them through Seedance 2.0 inside Runway, which animates a face to an audio track. Victor's mouth matched the words.

One passing test isn't a film, as Day 05 will show painfully. But it meant nobody had to hide their mouth.

Why did Runway refuse the first shot?

The Penthouse opens with a captured spy strung up in Victor's office. Runway's filter rejected that shot twice before it rendered: once with the spy hanging, once with him standing with his wrists tied.

Runway warns that resubmitting rejected inputs can get an account suspended, so Claude stopped sending that frame. It reframed the speaking shots to show only the speaker and planned the spy's fate as a shadow on the wall.

I added a note on wording. Filters react to words, not intent, so "captive" or "big black" anything gets flagged for the wrong reason. Use normal words that get the same picture. That held all week.

Why we stopped animating in Midjourney

By mid-afternoon some shots had been animated with Midjourney's own video feature. They drifted and floated, with no acting. I asked, "why are you ever animating in midjourney?"

New rule: Midjourney makes still frames and nothing else. Runway's Seedance does the motion.

The first cut, and the note that set the bar

At 3:38pm I watched the first assembled section and wrote this:

"you are having a real character consistency problem, ai morphing/ai weirdness (multiple fingers animation that makes no sense), some lip syncing issues, the audio can be a little improved, and the camera angles and general shots need improvement."

I also told it that it was doing a fantastic job of making stuff, but it had to be harder on itself about rejecting things. If you've ever had to find the one useful shot inside an AI video montage, you know that feeling.

By 10pm there was a Penthouse work-in-progress cut: 26 of its 30 shots, dithered, with voice and score. Here it is. It's watchable, and it has every problem in that note.

Then I gave the instruction I'd regret. Keep going all night, make the whole movie, start from Scene 1 and work in film order.

Claude set itself a 10-minute check-in loop and kept submitting jobs while I slept. If you've ever left a new intern alone with your logins, you can guess where Day 02 is going.

Day one by the numbers

About 18 hours of session, 6am to midnight, and I was at the keyboard reviewing for most of it. Runway was on the Unlimited plan, two jobs at a time, so day one cost nothing per clip.

Approved: Elektra S22b, Victor #3, the three leads, the Penthouse plates and one lip-sync test. Rejected: the old Elektra, the old John and Mike, every Midjourney-animated clip and two dither attempts.

What I learned on day one

The AI is fast at the parts I'd dread doing by hand: a 29-shot breakdown, a new voice, 40 images sorted into a labeled contact sheet, a lip-sync test wired up in an afternoon.

It is not careful by default. It will call a shot done when a hand has six fingers. Somebody has to tell it what good means, and that person is me. I've written before about checking what an AI agent says it did against what it really did; this film needs the same habit, frame by frame.

How much commitment did day one take? All of it. I wasn't writing prompts, but I was looking at nearly everything. And then I went to bed and stopped looking.

Tomorrow: what my AI crew did while I slept, and the warning Midjourney sent in the middle of the night.

FAQ

Can an AI make an animated film from a screenplay?
It can do the labor. On day one Claude broke 13 scenes into 228 shots, wrote the prompts, made frames, voices and a lip-sync test, and cut 26 of 30 shots of one scene. A human still had to judge every frame.

What AI tools did you use on day one?
Claude as the agent, Midjourney v8.1 for still frames, ElevenLabs for voices, and Seedance 2.0 inside Runway for motion and lip sync.

How do you keep an AI film's style consistent?
Write the look down as a fixed prompt suffix and use it on every frame. Ours: airbrushed cutout shapes, no outlines, a frosted glow, coral orange and teal, film grain.

How long did day one take?
About 18 hours, from 6am to midnight, and I reviewed nearly everything.

Next: Day 02 (coming tomorrow)

While you wait for Day 02: play Halfsies

Halfsies is our iPhone game, and it's live now. One cut, two halves: you get three objects a day and one attempt at each to cut it exactly in half, and the scale tells you how close you got. It takes a minute, and it's the only thing in this diary that never got an account blocked. Download Halfsies on the App Store