You Can’t Break a Rule You Never Learned
Why the AI feed is louder than it is useful, and what to do instead.
My feeds are full of the same breathlessness.
Nobody is talking about this.
This skill changed everything.
If you’re still doing it this way, you’re doing it wrong.
The implication is that there’s a switch somewhere, the person posting found it, you didn’t, and that’s why your week feels heavy.
This used to slide past me. It annoys me now.
It annoys me because underneath the noise, something real is happening. People are figuring out how to work with these tools. They’re learning what to trust, what to throw away, when to push, when to wait.
That’s a serious skill being built in public. It deserves to be traded the way scientists trade notes. Here’s what I tried. Here’s what surprised me. Here’s the part I still don’t understand.
Instead we’re getting miracle cures. Bottled. Sold. Forty-seven of them in a carousel.
Miracle cures don’t travel
Imagine someone wins the lottery and posts their numbers. You write them down. Next week you play those exact numbers, and of course nothing happens. The numbers didn’t win because they were those numbers. They won because they happened to match a draw on a specific day, with a specific machine, in a specific room.
You can’t borrow the win. You can only borrow the digits.
The prompts and skills flying past you in the feed work the same way. Whatever someone posted came out of their setup. Their memory store. Their codebase. Their prompt history. The workflow feeding into the call. LLMs are non-deterministic, and most of what makes a prompt actually work lives outside the prompt itself. It lives in the context that shaped the call.
The artifact is the part that’s visible. The invisible part is the fit between the artifact and the rest of your setup. That’s the part you have to earn for yourself. No one can hand it to you, because no one was inside your environment when they noticed what worked in theirs.
That earning is the noticing. It can’t be skipped, because there’s no artifact that contains it.
Even the model does this to you
The stranger version is that the model itself hands you the same kind of borrowed ticket.
I recently asked a model to write a configuration for itself, instructions for how it should approach a particular kind of task. It came back careful and prescriptive, twice the length I thought it needed to be. I’d written a few of these by hand and noticed that less is more with the current generation. Like a junior employee who has leveled up, today’s models need more why and less how.
So why was the model writing itself a long version?
Because it was trained on data from the previous generation, which needed the scaffolding. The model answering me doesn’t need it anymore. It just doesn’t know that yet. Its self-model is a model behind.
Reading the model’s advice about itself is like reading starlight. The light from a star can take years to reach you. By the time it arrives, the star itself has moved on. You’re seeing where it was, not where it is.
The model’s description of itself works the same way. The thing answering you is no longer the thing it’s describing, and it can’t quite see where it’s standing.
You can only feel that gap by playing with the current thing directly.
Trade notes, not cures
What we’re developing is tacit knowledge, the kind that doesn’t fit in a writeup because it isn’t made of words. It’s the cook who knows the pan is hot enough by the sound, the kind of thing that lives in your hands after enough hours that you stop thinking about it.
It’s always been like this.
Robert Rodriguez shot El Mariachi in 1991 for $7,000 in ten days. The way the story usually gets told, he hacked filmmaking. Showed up without a lighting person. Took long takes to save film. Edited it himself. That was supposed to be the secret to making a movie for $7K and selling it for over a million.
What’s missed in the telling is the decade before he got to Mexico with that camera. He spent it in his parents’ house editing on two VCRs. Dubbing tape to tape. Every cut a manual decision. Every mistake a do-over.
By the time he stood on set, he didn’t have a method. He had a sense for how scenes worked. The budget wasn’t the achievement. The achievement was that he could see what to do with the budget, because his hands already knew.
That’s what’s missing from the borrowed prompt. The prompt is a description of the thing. The thing itself only comes from hours with the current model on real work. Writing a prompt, running it, watching what comes back, adjusting, running it again. Most of it is unceremonious. Some of it is boring.
And when you do write it up, because you should, write it up like a scientist, not a salesman. Here’s the problem I was chasing. Here’s the prompt I landed on. Here’s why I think it worked. Here’s the part I still can’t explain.
That’s the trade worth making, and it’s the signal the feed is burying.
Get in there and play. Then tell us what you actually saw.