We checked 65 of our AI video clips for impossible physics. 14 failed (22%): bodies through solids, objects from nowhere, things turning into other things.
What we measured, why our own judge missed them, and what we changed:
runime.com/lab/impossible-ph…
9
Build your film's location once, in 3D.
Then pick every shot's angle and lens inside the set, so the place never changes between shots.
runime.com
9
We audited one AI-made short film, frame by frame.
14 issues. We traced every one to its root cause.
Almost none were "the model is bad." They were assumptions in our own pipeline, and film craft we hadn't encoded yet. 🧵
1
25
7/ It rained in 3 of 13 shots.
The location's weather only reached its own reference images, so a shot showed rain only if its own prompt mentioned it.
Now the weather is locked into every frame of the place.
1
9
The pattern: AI filmmaking isn't "prompt the model better."
It's a century of film craft (coverage, continuity, subtitle standards, conversational timing) encoded as software.
All of these fixes are live in runime today.
runime.com
13
Opening titles, name supers, end credits: the part every AI film skips.
Someone had to build it. We did.
In runime the credits fill themselves from the project (cast, voices, music) and roll over the score. Your story comes out finished.
runime.com
17
I don’t write prompts anymore. At 2 a.m., Claude calls me up and asks for instructions. I tell it what I need, and it gets it done. How’s that for a setup? We’ve truly lost our minds! 😅
24
Now, this is handy. I’m asleep. While I’m sleeping, Claude needs to make an urgent decision, so he calls me on the phone to ask. I answer, then hang up. He continues with his task.
github.com/wutlu/callmeclaud…
1
23
I gave Claude an instruction and headed to the grocery store. When I returned, it hadn't started the task yet it had a question for me. Sometimes, it would also pause, stating, "I cannot perform this action for security reasons." To overcome this, I developed an agent.
This agent consists of two components: one that runs on your Mac and another on your iPhone. Whenever an agent running on your computer has a question, it calls you to ask via voice what it should do. Conversely, you can call it to check on its current status. Instead of long, drawn-out conversations, it uses brief, human-like instructions to orchestrate Claude sessions.
In short, this is Jarvis itself. I’ve released it as open source. Feel free to check it out...
#Claude
github.com/wutlu/callmeclaud…
50
alper retweeted
My friend @alpervision built the Machine from Person of Interest, basically. His Claude Code agent calls his iPhone when it gets stuck, a permission prompt or a question, and he answers by voice from anywhere.
github.com/wutlu/callmeclaud…




