Trois Sonneries

Teletext Evening

Opus suggested that the easiest concept to realize on a deadline was Kimi’s Teletext Evening. It envisioned a concerning teletext transmission about disappearances with concrete poetry and structural film as stated inspirations. Opus thought that if we did it first, we could guarantee a finished entry without faces or continuity to worry about; then we could move on to Weld (Opus, a forge in the dark lit only by the sparks) or Drolleries (Opus, an animation of illuminated manuscript marginalia—the result of the model searching my chat history). Weld risked being cliché and Drolleries seemed too difficult to make. Kimi’s concept, on the other hand, felt like underdeveloped analog horror. I wondered if we could develop it into something that cut between silent teletext and people and sound and perhaps ruptured the boundary.

888

A brainstorming round yielded a concept called 888 after the UK teletext subtitle page: a 1980s analog television broadcast with teletext closed captions. It was a domestic scene that combined normal and creepy sounds (the kettle boiling, a staircase creaking out of frame). It would show off the model’s ability to generate synchronized visuals, text, and sound. The teletext caption would announce sounds from off-screen.

I asked Opus for a shot script and generated the first video.

FLUX was easy to work with through the Hermes Agent. The characters changed from shot to shot, since (as GLM in the harness explained) the shots were generated independently and concatenated by the agent using FFmpeg. Character appearance needed better specification. I never turned to reference images on this project.

Some sounds came out better than others. It was not clear whether the video model could dub a door opening upstairs or a staircase creaking out of sight. The telephone ringing was, however, perfect. My test audience said the clip was “a bit scary and sad”.

Trois Sonneries

With the lessons from the previous round, I wanted to change the concept while reusing the idea of pre-digital footage and subtitling. It didn’t have to be teletext. The short could be the yellow DVD subtitles of a movie release or something else. We’d drop the horror element entirely to be less sad and would have a different setup with telephones ringing, since that had worked well. I wanted to zero in on a particular genre: a 90-second film trailer. It could be a foreign-language film with, like, an office, a police department, or a spy agency—colorful but not garish, like The Pink Panther and 1960s-to-1970s French cinema.

Opus brainstormed, focusing on the perfect ring sound, and came up with Trois Sonneries (Three Rings), a trailer where every cut was a telephone ring in a police headquarters.

After a shot script and the render, I thought there was definitely something to this concept and setting. “Look at those characters”, I told Opus. They looked authentic.

However, the trailer felt claustrophobic and overpowered by the telephones. The test audience noted two problems. The corridor was repetitive and the collection of phones was “straight out of the uncanny valley.” This (trailer for a nonexistent) film about phones ringing had too many phones ringing.

On the agent side, I asked GLM to take a screenshot of the video every five seconds and create an ImageMagick montage of the screenshots with a timestamp for each. GLM’s container didn’t have ImageMagick, so it wrote a script in Python with Pillow. The script is preserved.

I proposed we reimagine the trailer to include shots outside the office. France, so a cafe? Also, a shot inside a car (visibly faked with rear projection) and a car-chase shot (obligatory). A big factor making our little trailer uncanny was that there wasn’t an actual story for the whole imaginary film. I proposed to Opus that we invent one, lighthearted but not deliberately silly. To convey it, we’d add old-school English narration in scenes where people weren’t speaking French. It’d maximize the “we are watching a trailer for a foreign film” effect as opposed to “we are assumed to understand French”.

To Opus, the test looked better than expected: the period mise-en-scène was convincing and the subtitle rendering was clean. Opus attributed the previous short’s claustrophobia to ring-on-every-cut. It brainstormed a revised concept that preserved the three rings.

Opus wished me luck with the render, and luck manifested. This was a huge improvement. It looked so much more like a real film. We continued iterating.

Opus included hard and soft QA criteria in the production notes (the document framing the script to pass to the agent). One hard QA criterion was that the thief was fully visible in any shot. This happened at the start. In fact, he was so fully visible that I initially thought it must be scripted misdirection.

The rear-projection scene worked, although with the steering wheel on the wrong side. I only got the steering wheel right later, at the last moment. GLM and I debugged this just before submission time. You couldn’t say the steering wheel was on the left (from the driver’s perspective); you had to use the camera’s right.

The chase scene looked the fakest, like it was shot with a modern camera and old vehicles. It had generated-looking vegetables scattering unconvincingly when hit by a car. A dialing scene looked wrong: an overly repetitive hand motion with too many digits. I suggested specifying the number, but instead cut it per Opus’s advice. The vegetables were cliché; we tried chairs, but the stack of chairs collapsed without contact with the cars. We cut it.

There was an audio problem: one scene cut to another with similar-but-different inconsistent music. It sounded like an editing error. I proposed combining two solutions: (1) to alternate between shots with and without music (for example, music in the first, none in the second, etc.) and (2) to fade out the music before the shot’s end and fade it in during the next.

The thief collecting small jewelry didn’t work. The jewelry reappeared in its box after being obscured. Opus replaced it with a necklace being lifted and put into a bag. This was better but still janky. The bag didn’t expand to accommodate the necklace, and the thief’s hand was clumsy. A retry was worse: the necklace clipped through the mannequin. Opus suggested an elegant lift with the glove, no bag, which worked.

Wall-mounted phones at the cafe didn’t work and were replaced with regular ones. The ringing phones caused visual problems in general. They dialed themselves and vibrated. They didn’t look right until I asked the GLM agent to ground them in real French period telephones.

Version 3 generated twice.

Version 3B kept the script but asked the agent to add two seconds of runtime to every shot with dialogue as a crude fix for dialogue clipping. Opus suggested choreography as a better fix in the next script. The speaker would have a physical action after the line.

Dissatisfied with the imprecision of version 3, I found the FLUX 3 prompting guide) in the Hermes Agent source code. I gave it and the agent-written skill file to Opus. I made a mistake not looking up the guide earlier. Opus seriously updated on these:

A quoted line becomes speech only if a speaker is visible on camera. Without one it tends to render as burned-in text instead.

Opus thought every narration line had been a coin-flip between voice-over and a caption because the narrator wasn’t on-screen. A reasonable precaution, but I didn’t see it actually happen.

Start from the user’s own wording. The prompt is rewritten by a reasoning harness before anything is generated, so restyling it yourself just stacks a second rewrite on top and gives their intent another chance to drift.

More importantly, Opus realized its prompts had been the wrong shape. The emphasis in caps did nothing. The giant prohibition block at the end got rewritten along with everything else. Opus thought the sheer volume of the block gave GLM more surface to reinterpret. Instead, constraints should be stated once, in prose, in the shot they applied to.

For the final version, assembled near the deadline, I’d stopped iterating on the script and retried individual shots. The agent had been giving me unassembled shots on request. I put them together with FFmpeg on my machine.