The AI Built My App. Getting It Online Took Two More Weeks.
I'm not a developer. I used Claude to build a working web app in a few hours. Then I spent two weeks trying to ship it. Here's what I learned.
AI tools are brilliant at writing code. They're unreliable at deploying it. The two jobs require completely different environments, and the second one is still a human problem.
I'm not a developer. I want to be honest about that upfront.
I'm a former pastor. My family ran a flower farm until we sold it. I learned to build websites by necessity, not education. These days I help small businesses in Northeast Alabama get found online, and somewhere along the way I started building software tools because the tools I needed didn't exist and the developers I asked were too busy.
About a year ago, I started using Claude to write code. I'd describe what I needed in plain English and watch it build. Real logic, real interfaces, real data flowing through. It felt like someone handed me a power I hadn't earned and didn't fully understand.
The first time I hit the wall, I'd been at it for about three hours.
I had a working app on my screen. Forms that submitted. Data that saved. A dashboard that pulled numbers and displayed them exactly the way I'd asked. I showed it to Heather and she said it looked real. It was real. It just lived at localhost:3000, which is developer language for "your laptop only."
Then I said four words to the same AI that had just written all of it: now make it live.
Two weeks later
I'm not going to walk you through everything that went wrong. Some of it I still don't fully understand. What I can tell you is that the AI that had written the app with calm precision turned into something different the moment I asked it to deploy. It generated configuration files that looked right and weren't. It gave me commands that assumed a machine I didn't have. It described infrastructure that didn't match any real service. It announced three separate times that the deployment was complete. Nothing was deployed.
The code itself was never the problem. Every time I ran it locally it worked. The problem was everything between "this works on my screen" and "someone else can use it."
Servers. Domains. Environment variables. SSL certificates. Database connections. DNS records.
The AI could describe all of these. It could not actually manage them. It was like asking someone to give directions to a place they've only read about.
What the article got right
A few days ago I read a piece by the co-founder of a hosting startup that finally put language to what I'd experienced. He talked about the feedback loop.
When an AI writes code, it works inside a tight loop. Write something, run it, see the error, fix it, run it again. The whole environment is right there on one machine. Everything is visible. The AI can iterate toward correct because it can see what happened.
Deployment breaks that loop. The systems are remote. They're stateful, meaning what happens depends on what's already there. They're owned by someone else, a cloud provider, a domain registrar, a certificate authority, and they don't give you a clean error message when they do the wrong thing. The feedback might come back as a failed DNS propagation eighteen hours later. It might come back as a billing charge at the end of the month. By then you can't remember what you changed.
No feedback loop means the AI can't iterate toward correct. It's not that it gets dumber. It's that the environment that made it smart no longer exists.
What I do now
I learned the deployment side by doing it badly enough times that I stopped making the same mistakes. I know how to wire up a Vercel deployment, connect a domain, configure environment variables so they don't show up in the code, set up a database that doesn't disappear between sessions.
Most people building with AI tools don't need to learn all of that. They need it done once, correctly, by someone who already knows it.
That's what I do for people now. You bring the app your AI built. I put it somewhere real people can use it.
It's not glamorous work. But two weeks of my time down a deployment hole taught me it's real work, and most people building things with AI right now are going to hit the same wall I hit.
If you're sitting at localhost:3000 with something that works and no idea how to make it work for anyone else, let me know. That part I can handle.