Coding since '88. Building agentic teams, https://xitter.cf/t.co/hsxPWrrHwl & https://xitter.cf/t.co/AThQajI6pT. Founder @byte5ai. Freediver. Film enthusiast. Here for what AI does to software, UX and art.
Frankfurt, Germany
Joined May 2009
- Tweets2.4K
- Following694
- Followers517
- Likes1.6K
Our Claude meetup on forging skills in Frankfurt was so full last night that the chairs ran out. Every other place to sit was taken too.
Thanks for the sharp questions and great conversations. Next one in 3 weeks.
Slides (in German): byte5.ai/en/talks
ALT Packed room at the Claude meetup in Frankfurt, Germany on Oct 13, 2026
Two new AI toys, and I can't try either.
Meta's Muse: not available in my region.
OpenAI's Dots: not available in my region.
Global hype, but local roll-out. Well done.
One of my big pains with agentic coding:
splitting work across several Claude and Codex accounts and remote hosts. Things get messy fast.
Just found Kepler by @GitKraken.
It handles those sessions on my servers, using all my subscriptions: Claude Code, Codex, Agy.
It's still an early version, but the approach is right.
Could become my daily tool.
Hope they will push this product to 1.0 soon.
I finished a nine-minute animated short in Seedance 2.5. Not one reverse shot in it, and that was on purpose.
Video models have no spatial imagination. Neither does Claude when it writes the prompts. Ask for a reverse shot and you get garbage. Tell the agent the camera turned 180 degrees and it mirrors the scene, which is wrong in a different way.
I learned that on earlier projects and sidestepped it entirely this time. If you need reverse shots, the room has to exist before the shot does. Panorama or 360 anchors per location, or clay geometry in Blender plus style anchors for the look. Geometry from the model, look from the image, action from the prompt.
In my case the room was one anchor plate per location, and finding each one took a folder of rejects. The whole production came to 919 generations and 7 GB of output. Made alone in four weeks for the @higgsfield_ai Global Film Festival. Link in the first reply.
What breaks first in your reverse shots: room layout, character position, or camera direction?
"Make no mistakes" at the end of your prompt. Classic running joke, does nothing, right?
But let's briefly take it seriously: added to a Claude or ChatGPT prompt, it does shift the output. Just not toward fewer mistakes. More often toward more hedging, or more confident-sounding wrong answers. Pick your poison.
What "make no mistakes" correlates with in training data isn't careful error-checking, it might be prompt guides talking about this exact "trick". Meme recognition, not carefulness.
Use it, and the output just proves the meme right again. That's the actual self-fulfilling prophecy.
Same story with threats: Sergey Brin publicly said models perform better if you threaten them. A Wharton study tested it. No significant effect.
Be honest, have you ever tried either one? Know any other prompt slop like it?
ai-film-production just shipped v3.4, a free skill for Claude or ChatGPT. What's new for the actual filmmaking side:
It treats the video model like your crew, not an intern you dictate every move to. Tell it what happens in the scene and when, and it handles composition, cuts, body mechanics, material and light on its own.
New camera angles can be pulled out of a clip you already shot, no reshoot needed.
About 30 documented media looks, animation eras, ink, felt, paper-cut, marionette, described by how that medium actually moves, not slapped on as a filter.
Free, MIT licensed, 25 chapters of sourced production craft, install as a skill in Claude or ChatGPT.
Genuine question: Has anybody finished an AI short film using Seedance 2.5, MiniMax H3 or Kling 3.0 without visible slop?
No location drift.
No character drift.
No broken interactions.
No weird movement or physics.
Drop the one worth sharing.
What tool or workflow held up?
PERIHELION.
Three small creatures on a world of ice each face what they believe is the end of everything. It was the beginning.
My entry for the @higgsfield Global Film Festival.
If you prefer watching PERIHELION on YT, here's the link:
youtu.be/e8_sz6yJZEI?si=L7eo…
The Eliza effect is strong with me when working with Opus 5.
Mostly I just teach it words it really shouldn't know.
OpenAI shipped GPT Image 2.5 this week. ai-film-production, a free skill for Claude or ChatGPT, already knows how to use it.
GPT Image 2.5's precision-editing mode means fixing one wrong prop or one bad shadow in a reference image without regenerating the whole thing and rolling the dice on everything else changing too.
25 chapters of sourced production craft, MIT-licensed:
github.com/iret77/ai-film-pr…
Astra: "Yeah, I should really do something about that!"
Me: "And?"
Astra: "Oh, I should've started already. Doing it now!"
Me: "Did you start?"
Astra: "No! That was wrong of me!"
Me: "Yeah, and?"
Astra: "In a situation like this, I should act instead of talking."
Me: "Correct. Have you started now?"
Astra: "No, still haven't."
Me: "Why not?"
Astra: "There's no reason or excuse for that."
Me: "JUST START ALREADY!"
Astra: "Yes, doing it."
Me: "You're still not working!?"
Astra: "Correct."
OpenAI's promo video showed GPT-6 Astra booking a tennis court and building a 3D game in minutes. Welcome to the AGI era, they said.
Mine stopped at a captcha, told me to call the restaurant myself, built a 3D game with graphics a 3-year-old could draw, then the server gave up.
Claude just runs a film production. 🎬
That is the shortest way to describe ai-film-production, a free skill I have been building. 25 chapters of sourced film craft that turn the model into your crew, DoP to line producer. You stay the director.
Why did I build it? Because I needed it myself.
AI video is easy to generate and hard to direct. Ask a model for a two-character dialogue scene and it invents a third line nobody wrote, lets one character wander out of frame and puts the red shirt on the wrong person. The model is fine. Nobody locked those things. Better adjectives won't fix that, a production method will, and it has to be built around what these models do when you leave something unsaid.
What happens when you say "plan my short film":
1️⃣ it reads the craft baseline for your genre and asks for the story before the style
2️⃣ cast, locations and props get built as approved reference assets first
3️⃣ every shot exists as an approved still before any video, attached as a reference rather than a start frame
4️⃣ every prompt runs through the same seven steps: canon, routing, references, writing, lint, delivery, review
5️⃣ a take that comes back 90% right gets repaired with the platform's edit modes instead of rerolled
The rules come from the official Seedance and MiniMax prompt-writing guides, from the production breakdowns @higgsfield_ai published for their open-sourced $2M feature film, from practitioner protocols and from my own sessions. Each source is named in the repo. Four rules I did not expect:
✅ cut the heads off the full-body panels of a character sheet. The small, soft face on a wide panel is the one the model copies into wide shots, badly
✅ state the exact number of words a take contains, or the model improvises dialogue. It does not like silence
✅ write "no CGI" twelve times and you get CGI. A repeated ban works like an attractor
✅ an army is never a number. One sharp warrior in front, a dense crowd behind, silhouettes in the fog
Every rule is labeled as verified, plausible or marketing claim, with a tag that points back to where it came from. Where sources disagree, the disagreement stays in the text.
Independent project, MIT licensed, no affiliation with any platform. There is an upload-ready .skill package in the releases. Grab it, drop it into Claude or ChatGPT and tell it "plan my AI film":
github.com/iret77/ai-film-pr…
If you make AI films, tell me what you're building and what your pipeline still gets wrong that no prompt seems to fix. I collect those. The good ones become rules in the next version.