The reel pipeline of Reads in Motion starts with a step nobody sees. Before any image or voice exists, a model reads the essay and writes down what it thinks the piece is about: the summary, the moment worth building around, the line a reader would forward. Then it makes scenes out of that read, and the read itself gets thrown away.
This time I kept it. Katherine Lazarevich let me run her post through the pipeline and then publish what the machine decided her essay was, with her marking what it got right and what it missed.
The essay
I asked people in the Karo (Product with Attitude) community if anyone wanted to try a small experiment: run an essay through my reel pipeline, then ask its author whether the AI had actually understood it. Katherine was the first to say yes.
When Context Becomes a Cage
What the machine saw
No edits below. This is the model's own read of the piece, produced before any images or voice existed.
Summary
Katherine Lazarevich describes how she built a second brain in Obsidian and fed the AI the full context of her company — and how that turned into a cage. The model started treating every constraint she had given it as untouchable, and stopped asking whether those walls were walls at all.
So she built a process called Grill: it reads the whole record of decisions and rejected paths first, separates the hard constraints (money, a deadline, the law) from the ones she had simply repeated long enough, and only then asks up to five questions.
The point is not to make the AI ignore the box. She wants to see which walls she built herself.
Thesis
Feeding an AI your full context makes it sensible enough to defend every constraint you gave it, so Katherine Lazarevich built a process that separates the walls that are real from the ones she built herself.
Key moment
The constraint audit and its closing line: some walls are hard — money, a deadline, a legal condition — and others only look hard because she has repeated them long enough, until a preference becomes company doctrine simply by living in a tidy note called "Decision."
I do not need AI to ignore the box. I need it to help me see which walls are real and which ones I built myself.
Tone and arc
Tone it read: dry, precise, wryly understated — an operator thinking out loud.
Arc it read: it opens as a win, because context engineering stopped the AI from proposing plans built for somebody else's company.
It turns when that same context becomes a cage, with the model seeing the walls clearly and no longer asking whether they are walls. It resolves in Grill.
The lines it pulled
Give an AI enough context and it may start thinking smaller than you do.
A preference can turn into company doctrine with alarming speed when it lives in a tidy note called "Decision."
I do not need AI to ignore the box. I need it to help me see which walls are real and which ones I built myself.
The hook it chose
Chosen: "She feeds the AI every wall. It stops testing them."
It kept that one because of what it leaves unresolved — why more context makes a model think smaller, and what she built to fix it.
It wrote two others and dropped them:
"The AI keeps pitching plans built for Elon Musk." — open loop: what she fed it to stop that, and what the fix quietly cost her.
"Some of her walls are money. Some she built." — open loop: how she tells the two apart, which is the whole process.
How it decided the essay should sound
The read does not stop at meaning. The same pass casts the narrator and directs the delivery, so the machine is also making a claim about the author's register.
It cast Rachel, and said why: a reflective, dry, first-person operator essay, where Rachel's grounded American register carries the wry understatement without turning it into a lecture, with stability low enough to stay alive on the wall-sorting turn. It set her at stability 0.38, style 0.48, speed 0.92.
Then it shaped the five beats so the reel moves with the argument rather than sitting flat: a decisive push on the hook, a low lateral drift through the setup, a zoom-out reveal on the key moment with the voice dropping more intimate, a fast punch on the five questions, and a drift out to the end card.
One transition throughout, a dissolve. The music swells only on the key moment.
What Katherine Lazarevich says
Did it get the point of the piece right?
At a high level, yes. It understood the tension: enough context can make AI defend the box instead of questioning it.
But it makes Grill look mainly like constraint sorting. The process also stops AI from following the emotional direction of my question. I shouldn’t have to tell it whether I want support or criticism before it starts thinking. The thesis mentions five questions, but misses what they are doing: finding the question I should have asked.
So the read is correct, but tidier and simpler than the real process.
Is the key moment the one you would have picked?
Yes. This exercise is one of the most important parts because the cage isn’t only AI’s. It’s mine too. :)
I carry these constraints into my own decisions before I give the same box to AI. When people say “think outside the box,” for me that means not treating the current reality as the only possible one for a moment. And that’s what I’m trying to do together with AI.
What did it miss entirely?
It leaves out the research outside my vault and the check of what work AI and agents could take on, including possible automations. But for a short reel, I think that’s fair.
Anything it saw that you did not know was in there?
Not really. It didn’t show me a new idea, but seeing the essay in another form made me think more about its rhythm and structure, and about which moments I would bring forward.
Would you have opened with that hook?
Yes, I would have.
Does the voice it cast sound like the way you wrote it?
No. It sounds too solemn and dramatic. I would keep the voice reflective, but make it more curious, like someone thinking through the idea and wondering where it leads.
I think the reel stays in one register for too long, and that doesn’t work for a reflective essay. The voice should change with the thought. The Elon Musk line was a small joke, but the reel gives it the same weight as everything else. In my head, I read it with a laugh.
The tension can come from the music, something closer to “Hand Covers Bruise” from The Social Network, rather than from the narrator.
You didn’t ask, but I answered anyway ;)
What did seeing the reel change for you?
I love the idea of turning an essay into a reel. While I was answering the questions, I kept wanting to rewrite parts of the essay. Not because the model got them wrong. Seeing the story as a video showed me another way I could have told it.
For me, Grill feels like an interrogation. The models question me about my decisions and assumptions, and it can be uncomfortable. That’s why my cover shows a police interrogation room. I would start the reel there, move into the walls of the cage, and end with the freedom that comes from understanding which walls are real and finding the question I was afraid to ask out loud.
It’s a funny feedback loop. I could have written the essay this way, but I didn’t. Maybe next time I’ll make an image sequence first and see how the story moves before I write.
I also like the visual style. The reel picked up the same dark colours and mood as my cover, so it still feels like my piece.
The two takes
First take — before Katherine’s input
Second take — after Katherine’s review
What I take from this
The first take caught the essay’s central tension, but reduced Grill to a constraint audit and made every line solemn. Katherine’s review found what was missing: the five questions are meant to get past the framing of the prompt and find the question she should have asked. The Elon Musk line was also a joke, not another dramatic beat.
The second take fixes most of that. It starts in the interrogation room, makes the mood-following problem explicit, and lets the music carry the tension. But it still ends inside the cage. Katherine imagined the last beat as freedom after seeing which walls are real, and that never reaches the screen.
If you write something you would like to see read this way, send me a post. I will run it through the same pipeline and show you the machine's read before anything gets published, and the reel is yours whether or not we publish the conversation.
