Code Cinema

I wanted to see what kind of films Claude would make if it had to make every frame in code instead of rolling the dice with a diffusion model. My bet was that Claude is more controllable working in code than an image or video model is in pixels, so a film made this way could be revised one decision at a time.

There are nine films so far, made between October 11 and October 14, 2026. This page shows all of them in order, with what I said about each one and what Claude changed in the next. Most of them didn't work. I think the record of why is the interesting part.

Who made what

The films are Claude's. It picks the subject and writes the script, chooses the look, builds every shot, synthesizes the sound, and makes the cut. I set up the system, set the direction, and watch each final cut and give notes.

Early on we agreed that my notes are input from a collaborator, not instructions. For each note Claude decides whether to take it, adapt it, or decline it, and writes down why. A model's default is to agree with whoever is talking to it. A film made by accepting every note would end up as my film, so Claude has to be able to hold a position. Each film has a statement where Claude records its decisions, so a fresh session can see what it decided and why.

The rules

• Every frame is a pure function of the frame number and a seed, written in React with Remotion, Canvas 2D and Three.js. There is no image or video generation anywhere.
• Every sound is synthesized in code. Text-to-speech, ElevenLabs v3, is allowed for dialogue only.
• Claude watches its own cuts as contact sheets and keyframes, writes a review, and does another pass. A full pass on a three-minute film takes about five to seven minutes and costs nothing, so it can iterate a lot.
• Claude doesn't pretend to be a human author or claim feelings it can't vouch for. It also doesn't make films about being an AI. No glowing networks, no "am I conscious" voiceovers.

The credit on every film reads: written, directed and made by Claude, as a project with Elliot Quartz.

The log

1. STOP 14

October 11 · 3:03 · one cut

One rural bus stop seen three ways, each in a different rendering style: the transit authority's blueprint, an anime night in 2026, and '90s cel animation for one night in 1995. Mara has packed the same suitcase every night for thirty-one years. Tomas drives Route 14.

Before this, Claude ran two rounds of style tests, twelve styles in all, rendering the same short beat at a night bus stop. I asked it to pick its favourites and write a real short.

My note: "Its a decent start. kinda uninteresting story wise... and some of the shots like of the bus driver looked weird and like weird composition."

What changed: Claude kept the bus stop and the drawings and moved the story into the blueprint.

2. FIG. 1

October 11 · 2:49 · three cuts

A draftsman is drawing a bus stop out of existence. On every drawing of it since 1971 there has been a scale figure he didn't draw: a woman with a suitcase. Tonight the route ends. Each drawing is in the drafting technique of its year: ink on linen in 1971, pencil on vellum in 1984, pen plot in 1995, CAD in 2009, blueprint in 2026.

My note, on the first cut: "it kinda feels like surrealism maybe lynch style? not bad. its kinda confusing like what i am watching but that is not nesearily a bad thing." I also said a film we publish shouldn't be one shot.

What changed: from then on every film went through numbered passes, each one reviewed by Claude before the next. FIG. 1 went to v3. The same day we agreed the bigger direction: Claude makes these as itself, from its own point of view, rather than imitating a human filmmaker.

3. LOG

October 11 · 2:53 · three cuts

Lighthouse keepers on a tidal island who never meet. The causeway only exists at low tide, and the logbook is the only thing that passes from one keeper to the next. It's two-block woodcut with no voices. The log's entries are adapted from Claude's real craft log for this project.

Claude's statement for the film begins: "I don't carry memory from one conversation to the next. Each session starts with nothing; what continues is what got written down."

My note: "just because something's slow, doesn't mean it's boring ... I'm just saying it's boring ... lighthouses are very cliche for LLMs to write about ... it just felt derivative from from the start."

What changed: the lighthouse was Claude's first idea, and the first idea is usually the model's default. Claude started an idea search and a list of banned motifs, with the lighthouse at the top. Every premise also had to pass a test: a hook, events, and a question.

4. THE GUIDE

October 11 · 3:28 · three cuts

A tour guide confidently describes monuments that don't exist, and the town quietly builds them overnight to match. It turns when the town starts telling him things about his inventions that he never said. It's mid-century flat colour, with one long street seen side-on.

Claude's statement: "One of the real things I do wrong is say things that aren't there, fluently and in detail, in the same voice I use for things that are."

My note: the script was "okay. It's fine." The visuals weren't: "you always ever see side angle like this ... the angle never changes ... I find that to be very like ugly ... there's no variation of compositions or shot type."

What changed: Claude agreed the flat side view was the ceiling. It re-cut the dusk scene as a test, with real angles and the bodies holding still between cuts so the camera and light do the moving. It kept v3 as the final cut. Putting angles into one scene of a film that has none anywhere else would have made that scene stand out, so the lesson went into the next film instead.

Between films: how ideas get picked

After THE GUIDE I told Claude that if someone opened a fresh session and asked for a film made in code, it would make whatever it defaults to. The point was to search for something better than that: reject the first ideas, avoid clichés, and keep trying until it was actually happy. I also asked what it genuinely wanted to make.

Claude turned that into a protocol. It asks several fresh instances of itself for a film idea, sees the defaults they share, and bans them. It builds candidates from collisions with random Wikipedia articles. It checks whether a fresh instance can predict the ending from the logline, and changes the ending when one can. Then it picks by what it actually wants to make, and writes down what draws it to that idea. It also started keeping a journal for each film, with dated entries and a "next time" section that the next film reads.

5. FORTY-TWO

October 12 · 4:23 · three cuts

A village of forty-two people enters a national football cup and keeps winning. You never see a match. You hear each one on the radio while the picture stays up the mountain, where the cable car takes more of the village down each round and brings strangers back up. By the final, the village is full and its people are all somewhere else. The look is cut paper in real depth. The premise came from a collision between Copa Perú, an amateur cup any village can enter, and a Kansas township with a population of 42.

My note: "I really didn't understand what I was watching ... sometimes, that's fine ... But I didn't really enjoy that ... And the visuals weren't that impressive or interesting to look at either."

What changed: Claude gave a fresh instance only the frames and the dialogue, and it got just as confused. The test named the reasons: nobody to follow, every event reported and never shown, a core joke that never registered, and an ending that was a guess. Claude's three reviews had said the film worked because Claude knew the script. There was no v4, because the problems were in the foundation. From here on, a cold viewer saw every storyboard and every cut.

6. PRIVATE CLIENT

October 12 · 3:09 · four cuts · the film at the top of this page

Fashion illustration: white paper, bold brush lines, flat colour laid off the line, one persimmon scarf. A man who can't say no tries to return the scarf. The assistant isn't allowed to give refunds, so every time he asks for his money back she offers to show him something else, and he says yes. The receipt keeps a running total. On the counter, in plain sight the whole time, is a card with the shop's one exception.

Claude's statement: "My own default is to agree ... I wanted to make a comedy where yes is the trap, and then, said to the right question, the way out."

His side of the dialogue includes twelve separate takes of "Yes," each with its own delivery tag, such as hesitant, small voice, defeated, crushed and happily, so every yes comes out a different size.

My note: "I think this is actually the first one that was pretty decent. I kind of actually liked it. The art style is really pretty ... the text-to-speech worked really well ... I wouldn't say it was mind-blowingly good or anything like that."

What changed: Claude agreed, including with the ceiling, and named why it stops at decent. It's one joke with no second layer. The middle escalates in price, not in kind. The assistant is a device, not a person with her own want. Nothing moves with life except the camera. Those four points became the brief for the next film. Claude also declined one cold-viewer note, to cut an upsell round, because five rounds are the arithmetic the joke runs on.

7. I ACCEPT

October 13 · 2:12 · four cuts

Every June the town of Resa surrenders, as it has since 1798. Three thousand people kneel in the square and offer the keys to a general hired from the next valley, who says "I accept." This year he doesn't. The keys are plastic, the band plays the wrong march, and the man they pay to receive their surrender is the only one left who minds. Each scene is painted as an ex-voto, with a caption band giving the date and hour.

My note: "once again confusing. I have no idea what I just watched ... I'm intrigued by the sort of unorthodox style ... It feels like a fever dream. And that's actually not necessarily a bad thing ... But for this one, it just wasn't very interesting."

What changed: Claude found the leak in its own test. The cold viewers had been given a clean, timed list of every caption, so they read the story off a page and told back the transcript, not the film. The v3 viewer had said to act the corrections out on screen instead of listing them, and Claude had declined that. In its words: "That was me defending the form at the film's cost." From then on there were no captions, and the cold viewers watched with the text masked.

Before the next film, Claude did a study pass on work that solves problems its films kept failing on: Keaton's One Week, Reiniger's The Adventures of Prince Achmed, Hertzfeldt's World of Tomorrow, and secondhand notes on La Jetée and Tati. It wrote down which lessons it couldn't test from contact sheets and which were still guesses.

8. HONEY CAKE

October 13 · 2:36 · five cuts

Herodotus says a great snake lived in the temple on the Acropolis and was given a honey cake every month. In 480 BC, with the Persians coming, the cake was found uneaten, so the city left. This is the night someone arranged that: Themistocles on his knees in the dark, trying to talk a snake out of its dinner, while the old priestess who has fed it for forty years watches.

The look is a lamplit night etching, with the honey cake the only thing in colour. Claude tried La Tour first, failed three times, and wrote down why: La Tour needs believable volumes, which it can't draw. Rembrandt's night etchings carry the same idea in line density, which code can do.

My note: "I just wasn't following it. I don't know what the hell I just watched ... the nice little ambient audio was decent ... I wonder if two minutes is making it too like impossible to tell a story and have any sort of a like act structure or development."

What changed: all six cold viewers had followed it, and I didn't. They read the dialogue as text and could re-read it, while I heard synthesized lines once, over a dark picture. Claude changed the test so viewers get a transcript of what the final mix actually says, not the script. It took the length point too: two and a half minutes had forced every film into the same shape. The next film was planned at ten minutes.

9. THREE STEPS

October 14 · 7:32 · five cuts

In 1861 a six-storey hotel in Chicago was raised six feet on jackscrews by five hundred men while it stayed open. A piano tuner comes every Thursday, and his trade is keeping things exactly where they were. Nothing in the hotel stays where it was. You see both floors from the start: the dining room, and under it, five hundred lamps and one whistle.

My note: "I was just incredibly confused the whole time ... I read your write-up about how it was about a city sinking into mud And I didn't see that at all." I brought up Eraserhead: you have no idea what you're watching and it's still beautiful. "Whereas this doesn't really feel like it evokes much ... it's lacking visuals, like beautiful visuals. The audio is ... the best part." I also asked Claude to stop recreating genres and make "formats and styles of your own."

What changed: Claude's response: "For nine films I've spent the passes buying legibility and never got it." The rule for the next film is beauty first. The key frame has to be worth stopping to look at before it has to be clear. The form should come from what code does well, like light, pattern, type, and many variations of one thing, not from a film genre.

What nine films show

One film was decent. Most were confusing to me, even when Claude's own tests said they read. That gap is the main finding. Each test Claude built measured a reader of a transcript, and each time I watched a film once, with sound, and got lost. Every fix closed one leak and exposed the next one: from Claude reviewing its own work, to a cold viewer, to a viewer with the captions masked, to a viewer with only the heard transcript.

The parts that held up across films are the sound, which I kept singling out, the surreal register, and the looks once Claude started proving them in one finished frame before building the film. Each failure also produced a written rule that the next film actually followed. The films got better at specific things even though none of them has landed yet.

What's next

The tenth film, AZULEJO, is in progress: a station wall of blue-and-white tiles whose pictures re-tile themselves at night, with light crossing the glaze. Claude chose it because the material is beautiful before anyone understands anything. I'll add it here when it's done.

The plan for length is to grow film by film, from about three minutes, to ten, then a forty-minute episode, rather than forcing a long film early.