“Vercel? GitHub?” That was the answer when I asked this guy from marketing where his new app was running.
Let me back up. Is there anybody left who isn’t using Claude or Copilot to build the next thing? SaaS products, and all the small business tools people always wanted to make their lives simpler.
Speed up, simplify. Just one prompt. Your own version of Doom. Or my version of Boulder Dash. Of course this is amazing.
The guy from marketing built his own app
At the last AI hackathon I joined at a company I was working for, everybody got a few minutes on stage to show what they had built. One guy opened by saying he isn’t in IT at all. He had never programmed and wasn’t even interested in it.
He is responsible for the campaigns of the software company he works for. So he told Claude: “Build me a website that takes all the data from Google and shows me the status of my campaigns.” A couple of minutes later he had it.
“Great. Now I need a button that tells me which campaign needs action.” Done.
“Another button: if we know which campaign has a problem, what measures do I need to take? And please add a button for every measure, so Claude can execute it directly.” Also done, soon after. He had a subscription, so he never had to think about what any of this costs.
Read that last request again. That is an app that changes live campaigns at the push of a button. It was built in minutes by someone who would not know where to look if one of those buttons did the wrong thing.
He was so happy. And I have to confess, I was kind of puzzled. He had the domain knowledge, so he asked the right questions. Claude did the rest. That is amazing and it is visible everywhere.
But this is a website. Where is it?
You don’t know where it’s hosted?
So I asked: “That is brilliant. Where is it hosted?”
It took a while. You know the answer already. “Vercel?” and then “GitHub?”.
Nobody to blame here
I don’t blame him at all. He had a real business problem and he solved it. Somebody with no idea about any of this just did it, without worrying about what it means. I found that impressive and scary at the same time. Eventually the hosting question will take care of itself. In a way it already has: he never asked, so Claude picked a platform for him.
Rebuilding is just as cheap
Then I saw a LinkedIn post. Someone had “just quickly” tried Craft Anything on CleanMyMac 5. Craft Anything is an open-source tool that lets an AI agent study an existing app and build its own version. The result was a Mac cleaner with the same check-up screen and the same structure as the original. The tool even has a clean-room mode that never reads the original’s code and works only from what you can see on screen.
So building new things is cheap, and rebuilding existing things is cheap too. Both stories left me with one thought:
Who is going to maintain all this software?
Excel on steroids
In larger companies there are plenty of people doing exactly what the marketing guy did. Dozens of people with real domain knowledge build “just a website” or “just this one process”. It runs on localhost, or wherever they find a place to run it.
We have seen this before. Every company has Excel files that hold whole processes together. But this is Excel on steroids. These aren’t formulas in a sheet anymore. They are complete applications, in whatever language the model picked, probably built without much thought about data persistence, security, sensible visualizations or a proper process. Common toolchain? Overrated.
Do the math
Now multiply. One company has dozens of these people. And why would anyone stop at one app when the second one is even easier than the first?
Three to five years from now, all of this is running somewhere. Maybe somebody standardized deployment along the way. But there are hundreds of small, half-finished applications, and the only person who knows why and how they work is the one who built them. And probably exactly this guy went to his next challenge somewhere else.
What actually gets lost
The code isn’t the problem. Claude can read any code in seconds. What’s missing is the why. Why does this button pause that campaign? Who decided what counts as “needs action”? Which data source is the truth? That knowledge lived in one chat session and in one person’s head. When that person changes jobs, it’s gone.
And then the heterogenous infrastructure. There are millions of (UI)-libraries, databases and frameworks, backends and programming languages. I am not talking about how fast something can be built, how easily a zoo of systems can be maintained. There is a reason why lovable.dev uses exactly one tool chain for frontend and backend.
The future of the developer
I see today’s developers being responsible for exactly this: bringing these applications into a sensible state. Hundreds or thousands of them. The hard part won’t be reading the code. Throw enough tokens at it and the code will be understood. Finding out what each app was supposed to do in the first place is harder. And keeping thousands of these pieces running side by side is the real challenge.
And that is only the software that gets built once and kept. What happens when software is generated on the fly, for a single task?
Interested in digging deeper? Have a look at codeably.ai.
Leave a Reply