I built a system to edit my videos, and then spent about eight minutes proving it wrong.
I gave it two and a half minutes of raw footage and asked it to tighten the edit. Cut the dead air. Keep it moving. Land near ninety seconds. It came back at ninety-two seconds with sixteen cuts. Every instruction honored. Every target hit.
I watched it and the only thing I could think to say was that the cuts didn’t feel natural.
So I opened the same footage and recut it by hand. Eight minutes, four cuts, ninety-seven seconds. I kept more of the original footage than the machine did, using a quarter of the cuts. The longest stretch it left unbroken was thirteen seconds. Mine was seventy-four.
It took me longer than I’d like to admit to see what had actually happened. The machine didn’t fail. I did. I asked for less dead air. What I wanted was to not be interrupted. Those sound like the same sentence, and they pull in opposite directions, because the harder you strip the silence the more often you cut, and every cut is an interruption. It hit my instruction perfectly and missed my intent completely.
And I only learned what my intent was by watching the version that got it wrong.
The genies were a warning
I’ve been saying for a while that AI is not a genie. I meant it as a caution about expecting magic. I’ve come to think the metaphor is sharper than that, and sharper in a way that should worry you more.
Go back to the old wish stories, genies and otherwise. The wish is almost never refused. It gets granted. That’s the whole horror of them. You get precisely what you said, delivered with total fidelity, and it ruins you, because what you said and what you wanted were never the same thing. Midas asked that everything he touched turn to gold, and it did, and then he sat down to eat. Nobody in those stories is undone by disobedience. They’re undone by obedience.
That’s the failure mode almost nobody plans for. We prepare for the machine getting it wrong. Nobody prepares for it getting exactly what we said.
The version of this you can’t prompt your way out of
Here’s where the usual advice runs out.
The standard diagnosis is that you were vague. Be more specific, write a better prompt, give it more context. That advice is fine and it solves the easy cases. It doesn’t touch the hard one.
The most expensive version I see in other people’s companies is the support inbox. A founder points AI at it with a perfectly reasonable instruction: read the ticket, answer it, keep the tone friendly. And it does exactly that. Every ticket gets a fast, polite, well-formatted reply. A portion of those replies are confidently wrong, and they’re already sent, to customers, under the company’s name. Nobody wrote a lazy prompt. Nobody wished and walked away. The instruction simply never covered what to do when it didn’t know the answer, because who thinks to write that down. You don’t specify the thing you’ve never seen fail.
I have a quieter version of the same thing from my own desk, and it cost me a week.
My content pipeline produced five pieces aimed at coaches. Two had cleared a review step whose entire job is to catch work that doesn’t fit the business. Both came back clean and went into the schedule.
I had changed who this business sells to thirteen days earlier. I’d written it down. I just never told the part of the system that writes. I scrapped all five and redrafted three, and the real cost wasn’t the drafts, it was that for almost two weeks I had a machine producing confident, well-made work for people I’d stopped trying to reach.
Nothing malfunctioned in either case. The instructions were followed. In mine they were clear, specific, and thirteen days out of date, and no part of that system was built to notice the difference. A person can carry out your orders or question them. A machine only ever does the first.
That’s the genie problem with the easy answer removed. It isn’t what happens when you’re careless. It’s what happens when you’re careful.
Why it happens, which nobody tells you
Both stories have the same mechanism underneath them.
When you ask for something, you don’t hand over your want. You hand over a description of it. The description is a stand-in, and it’s always smaller than the thing it stands for, because most of what you want is stuff you’ve never had to say out loud. I’ve never had to explain that a video shouldn’t interrupt me. Nobody ever asked. It sat underneath the request the way a floor sits underneath the furniture.
A person filling that request stands on the same floor you do. They’ve watched videos. They know, without being told, that chopping something into sixteen pieces makes it feel restless. So they quietly correct for the gap between what you said and what you meant, and you never find out the gap was there.
A machine optimizes what you actually gave it. Faithfully, all the way down. It has no floor of its own, and it won’t flinch when your description runs out, because it can’t tell that your description has run out.
So it didn’t misunderstand you. It understood you exactly, and you had underspecified without knowing it. You always underspecify. That’s not sloppiness, it’s what wanting things is like. With people the gap gets silently absorbed. With machines it gets faithfully executed.
This is not a you problem, it’s the whole market
If this were just me, it would be a story about being bad at giving instructions.
Look at the companies with the most money to spend on getting it right. Uber went through its entire 2026 AI budget in four months and then put a ceiling on it. Gartner forecast in 2025 that more than forty percent of agentic AI projects will be cancelled by the end of 2027, and the causes they name are escalating costs, unclear business value and inadequate risk controls. Not the technology falling short. And the 2025 MIT NANDA study is the one that should stop you: only five percent of enterprise generative AI pilots produce measurable revenue. The other ninety-five percent show nothing on the P&L.
Now ask what ninety-five percent actually means. It doesn’t mean ninety-five percent of the tools broke. The tools work. It means ninety-five percent of those pilots did what they were told and produced nothing worth counting. Ninety-five percent of companies asked for something, got it, and found that having it changed nothing.
That’s the genie at industrial scale. Not failure. Compliance.
The other way to get it wrong
There’s a second mistake, and it’s the opposite shape.
Earlier this year I wrote that AI is a hammer, and I still think that’s right. A hammer is a real tool. It does a specific job. You have to learn to hold it, and most of us are holding this one for the first time, which is why so many people are gripping it by the head.
The problem is never the hammer. The problem is owning exactly one tool and swinging it at every job in the building. A founder buys seats for the whole company, tells everyone to use the chatbot, and calls that a strategy. Nothing moves, because buying a seat was never a decision about which job needed which tool.
The way I’ve put it before: when the tools aren’t directed and managed properly, it’s like having millions of really smart interns all doing stupid stuff. The intelligence is real. The direction is missing. A very large number of very smart interns with nobody assigning the work is not an asset, it’s an expense with good manners.
So: not a genie, and not one hammer for every job. What’s left is narrower and more useful. A specialized tool, pointed at a specific job, with a person still standing where the judgment lives.
I’d been saying a version of that for months before I worked out why it mattered. My approach has always been to treat it as a tool, because I was quick to say it’s not a genie. What you put in and the parameters you give it define what comes out. So I started treating each one like a specialist with a narrow band, the way you’d brief an expert you’d hired to do one thing.
The half of this that is not a warning
Everything above is a caution, and if I stopped there I’d be leaving out the part that actually changed my life.
I’m not a debater. When my son needed help preparing for debate, I felt the specific uselessness of a parent who wants to help and has nothing to offer. I couldn’t teach him cross-examination. I couldn’t coach a rebuttal.
The obvious answer was a real tutor, and I’d already ruled that out. Not at ten at night, not for one kid, not at a price I was going to pay. It had stopped being a thing I was trying to fix.
So I built one instead. I couldn’t be the coach, so I built the gym. I called it Cicero. It generates counter-arguments on any motion, holds him to the structure and the clock, and gives him criticism on clarity and logic instead of telling him he did great.
That’s the other failure of imagination, and it’s quieter than the first one. The loud mistake is expecting magic. The quiet mistake is pointing this thing only at work you already do, saving twenty minutes, and deciding it’s overhyped. Meanwhile the problems you gave up on years ago never make it into the conversation at all, because you stopped seeing them as problems. They became weather.
The truest sentence I have about any of it is one I said out loud before I understood it: for the first time in my life, the things I could imagine I could actually build. Notice what that doesn’t say. The machine doesn’t do the imagining. That’s still mine, it always was. What changed is the distance between imagining a thing and having one.
Now here’s the part I’d rather leave out.
Cicero works. He uses it. But getting it to work took far more of my time than the idea sounds like it should, and it works to a degree rather than well. Past that degree, the next move is judging which of two rebuttals is actually stronger and why, and that’s the thing I don’t have. I could build the room. I can’t referee what happens inside it. So it sits there, useful and unfinished, waiting on judgment I’d have to go borrow from an actual debater.
The video system is the same story in a different key. It works, and I still use it. What it doesn’t do is run without me. I can’t hand it footage and take back what it gives me, and I don’t expect a better model to fix that, because the thing I’m adding at the end isn’t horsepower.
I’ve written before that you don’t need to be an expert in the content to be an expert in the container. I still think that’s true, and these two are where I found its edge. Being expert in the container gets you to works. It doesn’t get you to done. The distance between those is your own time and your own judgment, and it’s a lot longer than anyone selling you the tool is going to say.
What to actually do with this
The instinct after a story like the video edit is to get better at asking. Write tighter specs. Say more up front.
Do that, and know it won’t save you, because you can’t say the part you don’t know you’re assuming. The wanting always outruns the words.
So the discipline isn’t better wishing. It’s staying in the room. Look at the output, not the status. A green check only claims that instructions were followed, and following instructions is exactly the thing that just went wrong. Ask the smaller question the report can’t answer for you: is this the thing I wanted, or the thing I described?
When those turn out to be different, you’ve just been handed something useful. Your own specification, executed without mercy, which is the one view of your thinking a person would have been too polite to give you.
Own the wanting. Rent everything else.
The systems worth building keep a person at that seam on purpose. That’s the work I do at KyberFive, mostly with founders who already bought the tools and are wondering why nothing moved. If that’s where you are, start a conversation.