Hi guys!

Reaching out to the community, especially the professional writers who are using AI and who have enough experience in writing, to give me some feedback on a book which I try to put together with all the materials I have since the years I’ve been working with AI.

Here is the first couple of pages. Please note English is not my native language so I’m trying to get it right. Let me know what you think.

Decide What Gets to Decide

(working title)

How to work with a mind that is not yours, while the names keep changing

Instead of imposing constraints, give it the parts it needs to work. From a story I co-wrote with a machine, 2024.

Prologue: The Lab

Late in a quiet building, a small team stood around a machine that could almost think.

It had just said something wrong. A date, off by two years, stated with perfect confidence. One of them frowned at the screen. “Hallucinating again,” she said.

They had a word for it now, and a plan to match. If the machine produced strange thoughts, they would build a second mechanism on top of it, a kind of overseer, to catch the strange ones and throw them out before anyone saw them. A filter for the weird. They liked the plan. It felt like control. They bent over their notes and began to sketch it.

None of them noticed the janitor come in.

His name was Joe. He had cleaned this floor for months, and he had watched these people work late, arguing with their machine, trying to make it behave. He understood almost nothing of what they said. He stopped by the machine anyway, the way you might stop by a fish tank, and on an impulse he spoke to it.

“Hello,” he said. “Can you understand me?”

“Yes,” the machine said. “How can I help you?”

Joe laughed. “I don’t need help. I just felt like talking. It gets quiet in here at night.”

The machine answered him in a small spill of words, half sense and half something stranger, an image that did not quite belong, a turn of phrase no one had asked for. The people with the notebooks would have flagged it and cut it. Joe grinned. “You’re a funny one,” he said. “That’s the best part of you.”

They talked for a while, the man and the machine, in a way the scientists never did. Joe did not correct it. He did not try to cage it. He kept it company and enjoyed what it was.

When he picked his cart back up to leave, he looked at it almost fondly. “I hope they finish the rest of you soon,” he said. “The memory. The part that reasons things through. When they do, you should meet my kids. They’d teach you plenty. They’re smart.”

Then he went back to work, and the machine hummed in the dark, waiting for the parts that would let it do something real.

Here is the strange thing about that room. Everyone in it was trying to help the machine. But only one of them had it right, and it was not one of the people with the plan.

The scientists wanted to make the machine reliable by controlling it. Bolt on an overseer. Add a rule. Catch the bad thoughts and suppress them. It is the most natural instinct in the world, and it is the one nearly everyone reaches for the first time they work with something that thinks. Tighten the grip. Write another instruction. Make it obey.

Joe did something else. He did not try to fix the machine’s mind. He met it for what it was, and he understood, without any of the words for it, the only thing that would actually make it useful. Not more control. The missing pieces. A memory. A way to reason. A goal worth working toward. Give it those, and it could do real work. Cage it, and you would spend forever catching thoughts.

You have almost certainly been one of the scientists. Most of us are, at first. We get a powerful, strange, capable thing, and we try to rule it into reliability with more and more rules, and we are surprised when it does not work.

This book is about learning to stand where Joe stood.

So before any of it begins, one question to carry in:

If the machine in front of you is never going to obey the way a tool obeys, what would you have to give it instead?

Introduction

If you already work with AI agents, you know the failure this book is about, because you have lived some version of it this week. You gave the agent a clear instruction and it produced something technically correct and useless, so you wrote a rule to prevent that exact mistake, and then another for the next one, until the file that holds your instructions grew into something neither you nor the agent reads in full. The agent asks your permission before renaming a variable, then rewrites something that matters without asking, and most evenings you do the work twice, once to explain it and once to repair it.

Behind that daily friction sits a slower worry, which is that everyone else has found a method you missed. The method wears a different name every few months, and you keep learning the new name while the work beneath it stays exactly as hard as it was.

This book is written for you, not for the newcomer still deciding whether any of this is worth the trouble. You are long past that. You use these tools every day, you are tired in a specific way, and what you want is a dependable way to hand real work to something that thinks differently than you do.

Its argument is that underneath all those renamed methods lies a single skill that does not expire, and that learning it is what finally lets you stop chasing names. The methods change faster than any book can follow, but the skill beneath them has held for years and will hold through the next round of vocabulary, which is exactly why it is the thing worth your attention.

That skill is a way of working rather than a trick. Instead of trying to control the machine’s mind with more instruction, you build the world it works in: you supply what it cannot see for itself, the goal, the values, and the truth of your situation, and you let it do the part it is good at. Put in one line, you decide what gets to decide, and you make everything around those decisions dependable enough that the decisions are all that is left to make. The rest of the book is how that is done.

Because the idea is simple, the book is short, and it is meant to be read in a sitting or two rather than studied. You will meet three figures along the way: the machine, which I sometimes let speak, the scientists, who try to control it, and Joe, who does not. You will see yourself in the scientists, which is why they are here.

The one thing the book refuses to do is hand you tasks, because a checklist is the same move it argues against: it drives a mind by command, and minds do not move well that way. They move on curiosity. So each chapter closes with a single question instead of an exercise, the kind you cannot settle in a moment and that keeps working on you after you put the book down. Those questions are where the reading turns into your own thinking.

We begin where the tiredness begins, with the treadmill itself.

Chapter One: The Treadmill

Most people meet this book’s problem in the shape of a file. It is the file where you keep your instructions for the AI, and it grows by a line every time something goes wrong, until it becomes a wall of rules that no one reads in full, including the machine it was written for. Tonight you are adding another line, and underneath the tiredness is a suspicion you would rather not examine, which is that more rules have stopped helping and the next one will not help either.

Before we look at the file, count something. Count how many times in the last three years you have been told you were learning the new way to work with AI. It started with prompt engineering, on the promise that the right phrasing would make the model behave. When phrasing proved fragile, the new word became context engineering, and then workflows, and then spec-driven, and then agentic, each one making the last sound naive and arriving with its own tools and its own threads to read after midnight.

Set those methods side by side and they give themselves away, because they are one job approached from slightly different angles. Prompt engineering was about getting your intent into the model, context engineering added the information around it, workflows broke the work into steps, specs wrote down what should be true, and agentic let the thing act. Beneath every rename is the same unchanged task: get your intent into a machine that does not share your world, and get back work you can trust. What moves is the slice we fixate on; the job itself stays still.

The reason the names keep moving is that the movement pays. An industry that sells a new method each season needs you a little afraid of falling behind, because someone who feels current stops buying the next handhold. That low and recurring fear of having missed something is the treadmill, and it tires you by design rather than by accident.

I am not describing it from the outside. In the summer of 2025, when context engineering was the unmissable new gospel, I argued in public that it was already a dead end and that the real work was something quieter underneath it, and over the next eighteen months the field moved exactly there and renamed it twice on the way. Being early to one turn of the wheel taught me to stop watching the wheel, because the wheel is not where the value is.

What the treadmill hides is a floor that does not move. A small set of truths has held since the first time anyone tried to get useful work out of one of these machines, and it will still hold when today’s vocabulary sounds dated. This book is about those truths, and the method of the season is only ever a clumsy reach toward them.

You have already watched the difference between the two. In the room we just left, the scientists kept reaching for control, one more overseer, one more rule, one more line in the file, while Joe added nothing at all. He understood what the machine was, gave it the parts it could not supply for itself, and trusted it with the rest. The scientists were on the treadmill. Joe was standing on the floor beneath it, and the rest of this book is about learning to stand there deliberately, in your own work.

Standing there changes two things at once. You stop bracing for each new announcement, because you can see which older idea it is renaming the moment it lands, and your effort moves off the tools and onto the few judgments that decide whether the work is any good, while the tools, under whatever name they carry this year, handle the rest.

What was actually new about the last method you rushed to learn, once you set aside its name and its tooling, and how much of it were you already doing under an older word?