I am building a new website. Except I’m not entirely sure I am building anything. Claude Code is building it. I am mostly sitting here pressing buttons.
Accept.
Send.
Agree.
Allow.
Open Terminal
Copy, paste, return.
Accept.
At some point, Claude decided we needed GitHub. I have GitHub. Apparently I now also have a new GitHub repository. Something called Astro is running somewhere in the background. Or Astra. I honestly don’t know.
Claude does. And this is perhaps the strangest thing about building with AI right now:
I have never had this much power to build something I understand this little.
My Substack - Some Assembly Required - has always been where I work things out in public. It’s where the ideas are longest, sharpest, and least edited for palatability.
When Building Meant Knowing
I’ve been building websites in one form or another for a very long time. There was always a relationship between what you could make and what you knew how to do. You learned HTML. Then CSS. You figured out FTP. You understood where the files lived. You understood domains. You understood hosting. You knew that if you changed this thing over here, something would probably change over there.
Later the tools became more sophisticated. Dreamweaver. WordPress. Squarespace. Webflow. Wix. They abstracted away increasing amounts of complexity, but I still understood the basic relationship between myself and the machine.
I was operating the tool. Even when the tool made something easier, I was still the one making it. That relationship feels like it is disappearing.
Today, I’m not really operating Claude Code. I’m telling it what I want. Then I’m authorizing it to do things I don’t entirely understand in order to make that thing happen. That is a fundamentally different relationship with technology. I’m not building the website anymore.
I’m authorizing its construction.
Control Without Comprehension
There is something slightly uncanny about this.
I haven’t completely lost control. I still have veto power. I can stop Claude. I can reject a request. I can tell it I don’t like something. I can delete the whole thing and start again. But veto power is not the same as control. I can stop an action without being able to explain it. I can authorize a result without understanding all the decisions that produced it.
We have always delegated execution. The mechanic understands the engine. The passenger doesn’t need to know how the plane flies. Someone else understands the complicated bit. But AI introduces a strange new variation. The system is not simply operating machinery on my behalf. It is increasingly deciding how the machinery should be operated.
AI is beginning to break that relationship. My capability is going up. My comprehension is going down. And somehow, I’m getting more done.
That feels important.
Because we’ve spent the last few years talking about AI primarily as a productivity technology. AI will help us write faster. Code faster. Research faster. Design faster. Automate repetitive tasks. Save time. All probably true.
But I’m beginning to think that’s not the most interesting transition happening.
The more profound transition may be psychological. We are learning to become comfortable authorizing actions we don’t completely understand.
The Button
The word changes. Accept. Allow. Approve. Authorize. Yes. Send. It’s always the same button.
Claude wants access to something. Allow.
Claude wants to install something. Yes.
Claude wants to create something somewhere. Send.
Claude wants me to authenticate with another service. Sure. Authorize.
Each individual decision feels trivial. Together, they’re something else entirely. They’re little acts of delegated authority. I’m essentially saying: I don’t know exactly what you’re doing, but I understand what I’m trying to accomplish, and I trust you to figure out the machinery.
Today it’s a personal website. That’s relatively harmless. But follow that trajectory forward.
Create the repository. Accept. Deploy the application. Confirm. Purchase the media. Approve. Negotiate with the supplier. Send. Change the pricing. Yes. Move the money. Authorize. Hire the candidate. Accept.
At some point, this button becomes very powerful indeed. And perhaps a dangerous one.
Not because the machine is necessarily doing anything nefarious. But because we are creating systems in which enormous capability can be exercised by people who increasingly don’t understand the mechanisms producing the outcome.
We Have Seen This Before
Of course, civilization already runs on abstraction.
You can drive a car without understanding fuel injection. You can board an airplane without knowing how to fly it. People send money across the world without understanding the banking infrastructure moving it. Someone can turn on a light without understanding the electrical grid.
Modern life would be impossible if comprehension were required for participation. We solved that problem by building layers of abstraction. Someone else understands the complicated bit. I don’t need to.
But AI introduces a strange new variation.
Because the “someone else” is increasingly not someone. It’s a system. And more importantly, the system isn’t simply operating machinery on my behalf. It is beginning to make decisions about how the machinery should be operated.
I tell Claude: Build me a website.
Claude decides what that sentence technically requires. What framework. What files. What dependencies. What architecture. What commands. What services need connecting.
I haven’t delegated the typing. I’ve delegated a significant portion of the decision-making.
That’s different.
From Operator to Governor
Maybe this means we need to rethink what technological competence looks like.
For decades, we have valued the operator. The person who knows the software. Knows the shortcuts. Knows the syntax. Knows which buttons to push. But increasingly the machine knows which buttons to push.
So what exactly is my job?
I suspect the answer is judgment. My job is not to understand every command Claude runs. It is to understand the goal, the boundaries, the consequences and the point at which I need to intervene.
Knowing what I want. Knowing what good looks like. Knowing when something feels wrong. Knowing which decisions deserve scrutiny. Knowing what can safely be delegated. Knowing when the machine has misunderstood the intention even if the execution is technically flawless.
And perhaps most importantly: knowing when not to press Accept. That is a very different kind of expertise. We may be moving from being operators of machines to being governors of machines. A governor doesn’t need to know how every road gets paved. But they damn well need to know where the roads are supposed to lead.
The button looks exactly the same at every stage. Only the consequences change.
Fuck It. Send.
There is a temptation to turn all of this into another warning about AI.
That’s how I feel. I’m uncomfortable, sure. But I’m also exhilarated. Because sitting on the other side of that discomfort is a real expansion of what one person can do.
There are things I can now make that previously would have required people with skills I don’t have. There are ideas I can test without assembling a team. There are technologies I can use without spending months learning them. There are entire categories of capability suddenly available to me simply because I can describe what I want.
That’s extraordinary.
So yes, there is a moment when Claude Code asks permission to do something and I stare at the screen thinking: I have absolutely no fucking idea what you’re doing.
And then I press:
Allow.
Maybe that’s reckless. Maybe that’s trust. Maybe that’s what technological progress has always felt like at the edge: a moment where our ability to use something arrives before our ability to completely understand what using it means.
For now, I’m going with it.
Because the website is getting built. The thing works. And somewhere behind my screen, agents are talking to systems, creating files, connecting infrastructure and doing things that would have sounded like witchcraft to me not very long ago.
I’m still in control. I think. I just don’t understand most of what I’m controlling anymore.
And increasingly, I wonder whether that’s the deal we’re all about to make.
So I Approve it. I’m still not entirely sure what I’m approving.
