I had an idea for an app. A friend showed me Lovable. I’m not a programmer. I don’t know how to code. In two months, I launched my first app. But first things first…

Using Lovable was a groundbreaking experience. Thanks to the text-to-code functionality, I saw the first prototype of my landing page. The euphoria ended after a few days. The available tokens disappeared after several minutes of working with the app. I had to wait until the next day to continue. My plan to develop the application on weekends ended as quickly as it began. The same friend sent me a YouTube video where someone talked about a more effective way to create code using text-to-code.

A breakthrough

After watching the video, it was clear that this alternative to Lovable application development process would not be as easy as before. However, my curiosity outweighed my concerns. I installed Visual Studio Code. I installed Codex. I opened a window with ChatGPT and… time has stopped… After 10 hours, I had a landing page, a local server set up, and the first draft of the next page with the beginnings of the application’s logic.

The first significant difference in this mode of code creation is the lack of time constraints. I built the application for as long as I wanted. The only limitation was the length of individual chats with ChatGPT – but more on that in a moment.

The second difference. In this mode of creation, I learned new terms used in the IT world. I had the impression that my specialized vocabulary was doubling every day. The euphoria returned. The development of the application was clearly noticeable. 

Engineer and System Architect

After several intensive sessions, I got the impression that other “people” were involved in the application development process. As strange as it may sound, when I sat down for another session, I felt like I was connecting remotely with a “team” that was supporting me. A two-person team. An engineer-coder and a system architect.

With the system architect (Chat GPT), I determined the best strategy for developing the application at a given moment and received prompts based on my ideas, requirements, and comments. I forwarded the finished prompts to the Engineer, who physically created the code, reported on completed tasks, and proposed the next stages of code evolution from the point of view of application development. 

The communication with each of them was different. Chat GPT was more “talkative,” proposing scenarios and pointing out the advantages and disadvantages of specific solutions. It reacted to the emotions contained in the communication. When another change caused regression, and I wrote that I was impatient and saw no progress, Chat GPT agreed and changed its approach. It no longer suggested another modification “blindly” but “considered” ways to diagnose the situation before the next code modification prompt. The chats resembled a creative conversation conducted in the form of a workshop. Sometimes it asked for feedback from Codex, evaluated and accepted it, or suggested another “patch” change that “sealed” what Codex had done. 

Engineer / Codex is very „economical” in communication, “focused” on performing a specific task. Communicating with him was like working with an expert who knows a lot and is capable, but does not share his knowledge as spontaneously as Chat GPT.

An emotional roller coaster

Late at night, another prompt, sending to Vercel, checking how it works – satisfaction. It works as planned – I can go to sleep. 

Second scenario. Similar time, another prompt, sending to Vercel – blank screen… everything fell apart, and the change was just adding another seemingly minor functionality change. I no longer have the strength or desire to analyze it – I need to go to sleep. But sleep doesn’t come quickly. How badly did it fall apart? What will work after the repair, what will stop working, and what will start working differently?

Another day, more prompts. Success, joy, mistakes, irritation… maybe start all over again? Or maybe I’ll try adding another diagnostic tool? And so on and so forth. It’s important not to get discouraged. Further experiences and getting out of situations that seemed hopeless helped to maintain optimism and the desire to make further improvements. 

Tired coworker

Communication with GPT chat and Codex takes place via chat. The longer a single chat is, the slower the communication. I had the impression that my “colleague” was feeling tired… This is especially true for GPT chat. Codex seems to be more resistant to long communication. The solution may be to close the chat and open a new one. This helps to return to a quick exchange in the chat, but… well, it always involves the risk of losing the thread from the previously closed chat. As long as we are in one chat, we can “count” on our system architect to remember what we “talked about” at the beginning of the chat. I feel that it helps a little to move related chats to one directory and open new ones in the same directory. What works is to remind important rules in the first prompt of a new chat. For example: “This is a continuation of work on the XYZ application, the limitations are Vercel’s policy rules and a limited number of functions, do not change the logic of the application, before introducing new functionality, ask about everything that will help build a consistent prompt without the risk of regression, etc.” This improved the quality of prompts created by Chat GPT for Codex in the new chat. 

Working with an impossible optimist

If the chat says that this is the final, precise, conclusive, definitive prompt – don’t believe it. Most often it isn’t – but one of the next ones will be. Don’t let that discourage you – “that’s how it communicates.” Don’t waste your time reading all the chat comments it gives after generating the prompt, such as: “Three reasons why this will work…” – it usually doesn’t work, it’s a waste of your time. 

Sometimes it’s easier to solve a problem with Codex, sometimes with GPT chat – experiment. You can make visual changes to the application faster in Codex chat by displaying the changes on a local server. The changes are visible immediately. Building the logic of the application and its structure is easier to prepare with GPT chat, which will do it more comprehensively.

Don’t be fooled by its optimism. Sometimes it will ask you to paste some feedback and comment: “Great, that’s 100% enough, that’s what we were looking for, now we’ll fix it once and for all…” and surprisingly often, this is only the first small step in the process of fixing or creating an application. 

Turning the Impossible Into Reality

We are receiving information from various sources about the threats and opportunities posed by the popularization and rapid growth of artificial intelligence. One of the undoubted benefits is that it allows people like me, who have no idea about creating applications or coding, to do so, and today I can do it and derive a lot of joy and satisfaction from it. And what are you doing today thanks to AI that was beyond your reach not so long ago…?