DE

Workshop Film Finished

SCARED REDUNDANT

Who gets scared by a machine that is never unsure?

1:56 English Fully AI-generated. Characters and ride are fictional.

From the workshop

script versions until the metaphor held
4
from the first line to the finished film, one night included
≈ 22 h
shots in the film, 30 clips generated
22
my time: 2:29 h measured on my turn, plus 50 % viewing time
3:43 h
actual AI cost: image, video, sound and Claude subscription share
≈ €25
the same Claude tokens at the API list price
≈ €75

What it cost

Image and video (Higgsfield) 554 credits: character sheets, sets and start frames (224 credits), 30 clips (330 credits) ≈ €23
Voice and sound effects (ElevenLabs) unit voice in the subscription, scare sound in four variants < €1
Claude Code as directing assistant share of the Max subscription; 284 million tokens, 549 model calls ≈ €2
Music (Suno) my own Suno subscription, no unit price in subscription
Total≈ €25

The idea

It started as a Halloween episode: a ghost train at a small English funfair whose ghosts are on the payroll. The owner wants to replace them with animatronics. The first animatic held up until the robot’s first test. After that it went flat.

So I rebuilt it. The boss went, and so did the supporting cast. What remained is a staff of three: a monster who has been frightening people since 1818, a young woman who learned scaring by watching, and a bride who never says a word. Then come three animatronics, all with the same scream.

The staff’s objections are the objections to AI. Nobody in the film says “AI”. Nobody says “ghost”. The joke comes from the dry tone of an office mockumentary. Four script versions in one day, until the metaphor held.

A scene is one scene

The first set of images looked good and still didn’t add up. At one point the grim reaper stood there twice, the robots were sometimes on the track and sometimes against the wall, and the bride jumped across the room between two shots. Every image had been invented from scratch.

The fix was a floor plan before any image: a straight corridor, the track on the right, a platform on the left, fixed positions for every character and fixed camera sides per scene. After that, every continuous scene came from a single master frame. Only what changes gets changed. Closer shots are crops of it. So the bride stands where she stood before.

Ghost-train corridor with three animatronics on a platform, a young woman and the monster in front of them, the bride against the wall on the right.
Drag: on the left the 3D blocking of the final scene, on the right the start frame for the clip.

What the machine couldn’t do

The video model moves everything that looks like a character. The robots bent at the hip, the witch took a step, the grim reaper started walking in the background like a person. Two attempts with very mechanical directions changed little. It isn’t a prompt problem. The model only knows puppets as people.

The fix came from the director’s chair: the robots don’t move at all. The model got them as switched-off shop-window dummies, rigid. The scare happens in post, only through the eyes and the sound. The LED eyes are off and flare hard red at the scare. Add a servo clunk and a canned scream from a small loudspeaker, the same file for all three.

Only then did they look like robots. Little movement, a cheap signal, always the same. That is also the core of the film.

The sound

The units speak with an assistant voice from the ElevenLabs library: “Hi! I’m here to help!” Voice and scream run through a small loudspeaker and the reverb of the corridor. Their level follows the rattle of the cart. That puts them in the same room as everything else.

The clips bring their own room tone, and it breaks at every cut. So each shot got a short crossfade, and a continuous room tone runs under the interviews. An audience laugh the model had put at the end of one clip is gone.

What it cost

554 Higgsfield credits for character sheets, sets, start frames and 30 clips, about 23 euros. Voice and sound effects came to less than one euro. The directing assistant ran on my Claude subscription: 549 model calls, 284 million tokens, about two euros as a share. Billed through the API, it would have been about 75 euros. I was on my turn for two hours and 29 minutes, just under four hours with viewing time.

About 40 percent of the credits are not in the film: images from the first three script versions, images from before the floor plan, clips with robots that moved too much.

What I take away

Space first, then the image. Without a floor plan, the model reinvents the room in every image. And a scene comes from one image, not from ten.

What a machine can’t do can often be left out rather than forced. The robots became believable once they were allowed to do less.

What this means for an organization

Strategy. The question in the film is the same as in any organization: what should the machine take over, and what not? The units scare reliably, simultaneously and always the same way. That is exactly why they scare nobody. The answer lies in the purpose, not in the tool.

Role. Someone has to define the space before the machine starts. Without a floor plan, the model reinvented the scene in every image. Whoever sets the frame saves rework.

Process. Derived, not regenerated. Every scene came from one master frame, and only what changed was changed. A tendency of the machine that shows up twice is not paid for a third time. That is when the method changes.

With the human. The judgment of when something is funny. The timing of the cut, the small flinch instead of the big reaction, “Holy shit.” as the last line. The machine delivered thirty clips. Which moment carries was my call.

Division of labor

Who did what?

Human

  • The idea and the first script, a Halloween episode in a ghost train
  • The rebuild into a metaphor: animatronics arrive, the staff have doubts
  • The title SCARED REDUNDANT and the unveiling gag “Hi! I'm here to help!”
  • Every note on space, axes and positions, until a scene was one scene
  • The decision that the robots do not move at all
  • The ending: robot tips, everyone flinches, the bride is startled, then “Holy shit.”
  • The music, made with Suno, and every listening judgment

Machine

  • Script versions, shot list and floor plan (Claude as directing assistant)
  • 3D blocking, character sheets, sets and start frames
  • Scenes as a chain from one master frame, so positions hold
  • 30 clips with native voices, checked by speech recognition
  • Eye flash, loudspeaker sound, room tone, edit, titles and cinema look by script

Tools

  • Claude Code (Opus)
  • Higgsfield
  • GPT Image 2.5
  • Nano Banana
  • Kling 3.0
  • ElevenLabs
  • Suno
  • Whisper
  • ffmpeg

Notes

Fully AI-generated. Characters and ride are fictional.
Idea, script, direction, music and sign-off: Stefan Margenfeld. All characters and the ride are fictional. Voice of the units: ElevenLabs library voice “Ava”. Music made with Suno. Higgsfield at list price (1,000 credits = 49 US dollars), ElevenLabs at the stated price per generation. Claude Code as a share of the Max subscription (200 US dollars a month), calibrated on the measured weekly quota; comparison value at the API list price of Claude Opus 5.5. Rate €0.86 per US dollar. My time: measured time on my turn plus 50 % viewing time.

How I measure →

Keep thinking · From the journal

What future do we want to build together?

Let’s talk. Next workshop notes: THE AGENCY