I just used AI to build a custom application, and now there is no going back.

The Process

What my friend had been doing, I thought, was quite a novel approach. He would instruct one LLM to generate a prompt, to get Claude Code to build the application code. The general description of the prompt was fascinating to me because the LLM would fill in gaps of thought and since you could see the prompt you could immediately see if what it was going to do fit your objective. As an example, when my friend was demonstrating this flow to me, he asked the AI to build a web page for a specific purpose, but didn’t list all of the exact pieces of information that the page should present. The AI *understood* the theme and added details that were not specified, but that made sense. In that flow, a person can look at the generated prompt and tune their idea, deciding if those fields were necessary. Or if it was highlighting a larger conceptual issue that needed rework. Without ever even writing any code. It’s like, you tell it what you want in plain English, and it then writes a technical specification.

When I sat down with ChatGPT, I had planned to do it the same way. I told it from the start that my goal was for it to generate the prompts and I would use Claude to write the code. My thinking was that I would build it modularly and in participation with the AI with my active input in IDE.

I started my process by describing to ChatGPT what I wanted to do, and uploaded a screenshot of a spreadsheet that I have been using as a rough idea of the problem I was trying to solve.

We went back and forth and eventually it started generating mock images, and then HTML prototypes inline with our chat. As the prototype became more sophisticated, it suggested that we start mocking it up using real HTML and JavaScript.

Within a couple of days we had the prototype looking quite nice. We were working on my local MacBook and I could run the prototype in a regular web browser. We got it to the point that I wanted to see the prototype on my phone. By this point I’d developed quite a bit of trust over what the tool could do, but I hadn’t even scratched the surface of its capabilities yet. I asked, “Hey, this is nice being able to see it in Safari on my Mac, but how would I view this on my phone?”

Within a few minutes, ChatGPT had installed a web server, updated my firewall settings and then told me what address to paste into my phone’s browser to see it.

Amaze, amaze, amaze.

I didn’t lift a finger. I just sat there sipping my coffee while it went off and did all of the work.

At this point, our design loops went into high gear. The prototype ran so well, and was so easy to adapt by just telling it what I wanted to see, that I began to realize that there may be no need to go down my initial path of using Claude to write the code. The prototype was already the front end code.

I asked it a question about what it had built and it explained to me that the prototype was written in JavaScript and that it had decided to hard code all of the HTML, JavaScript and the sample data into a single file for every version that we would iterate on. That way changes would be progressive, and nothing would be forgotten between changes.

So I asked, “Would you be able to write the production application?”

It answered confidently and said that the production code wouldn’t be organized this way. That it would break the work into the types of files and structure that a typical enterprise application would use.

So I asked, “Where do we start?”

Within an hour, it had built out a project connected to my GitHub account. It walked me through fixing my SSH authentication so that my new MacBook could talk to GitHub. It did a custom uninstall for some old tools that I didn’t even know were still on my machine. It installed HomeBrew, Xcode Command Line tools… all of the things that a professional developer would do. While I just sat back and sipped my bourbon.

I’ve been using ChatGPT for a few years now for a variety of things, but I’d never used it this way. I’ve never had it interacting with my desktop. There were a number of permissions that it needed to be able to do this, but once it had access, what it is capable of is kind of mind blowing.

The AI and I went back and forth for a few days making the prototype richer and richer.

Early work was using the default model, which is great and fast, but it’s expensive. After ChatGPT burned through my allotment of credits rather quickly on something that I asked it to do, I went out to the public interface and asked it questions about the credit usage. The AI suggested that writing code isn’t that complicated and that the model that I was using (ChatGPT 6 Astra) was overkill. That my credits would go much farther if I switched to a smaller model.

Of course it was right.

After switching to ChatGPT – 6 Sol ‘Medium’, it just sips credits now. I’ve only run out a couple of times in a five-hour window since then. By that point I’m usually ready to live my real life for a while.

After getting the project off of the ground, with it pushing code to GitHub I took things to the next level. The production application would need to be hosted on a commercial server in the public space so that it could be used like any website on the Internet. I didn’t want anything to do with passwords, but it would need authentication.

I already have a couple of websites, and one that is “chiseled by hand” is using a hosting service that I’m familiar with. It seemed like using Google as an authentication system shouldn’t be that hard. And, since it seemed like this AI was a wizard at all things, i asked it:

“Hey, what do I need to do to get this site hosted on the web using my service provider with Google authentication?”

Within an hour, it had gone to my hosting services website, read their documentation about what tools and technologies they supported for my tier of service, and decided to rewrite the application in a different language, because my service could host that one without me going to a higher tier!!! It did *everything*. I was opening browser tabs, going through the online forms, filling out everything that was needed like configuring the database, setting up SSL — the only times it asked for my input was when it wanted me to come up with passwords for things that it didn’t want to see.

It went into my Google Cloud console and set up what it needed for me to have Google’s OAuth authentication system as a front end to my site.

It was truly humbling.

I’ve done by hand just about everything that it was doing at some point in my life, and I’ve mostly enjoyed doing it. I mean, I love(d) writing the code, but going through reference documentation and the tedium of wiring up all of this crap is never any fun. Sitting back and watching the AI do what would have taken me days of banging my head, going to stack overflow, googling solutions to esoteric error messages… there was none of that this time. It just took care of all of it.

After this experience I feel like I have a better sense for what it can mean to build an application using an LLM to write the code. I started designing this project with ChatGPT on September 17th. By around September 25, a week later… I had progressed beyond prototype to working demo using code and a local database. By October 2nd, I had a public web address and a working app on the internet. Two weeks. And never once have I written a single line of code. I’ve browsed the source code out of curiosity to see the types of decisions it made about naming files and decomposing logic, but I didn’t go any further.

One of the most impressive things to me as a person who worked in this space as a professional for almost half of my life is how the AI covers gaps and edge cases. For example, in the real world, one frustration that developers are often met with is a design that is presented as complete but it only shows what happens when things are successful —no screens designed for what happens if things go sideways. It burns a lot of cycles going back and forth between the product team, design team and development teams. But when you give AI a description of what you want without those things clarified, it just makes choices for you. I didn’t always love the choices it made, but not once did I ever see the code that the LLM generated crash. It always handled the unexpected in a way that was reasonable. I could then go back and give it more specific guidance as to how I wanted to handle the situation. When a tester entered a dollar value that was bigger than an integer could represent, as the tester (my lovely wife) did, the app handled it. It gave her a clear error message about what had happened. I later went back and told the AI not to let users even enter values that large which is a better user experience, but the way it initially decided to handle it wasn’t wrong. And it didn’t crash!

There was one situation where I asked the LLM to “close dialogs when the user deselects a drop down menu”. Instead of just doing exactly what I asked, it pushed back. It asked me, “Do you really mean close the dialog when they deselect the dropdown, or when they deselect the dialog.” I was very impressed. My sloppy language was not what I wanted, it was not reasonable. The AI realized that, pushed back and asked if I was sure.

One thing that has become even more clear to me than ever before is that an application written by AI, isn’t developed by AI. The intent comes from the operator. The intent is a human desire. What AI brings to the table is the ability to turn your idea into reality at speeds that not even the best coder on earth could come close to.