I’ve been working with AI coding agents for a while now in my professional environment, and overall, they’ve proven to be quite effective. While they occasionally fail or lead to inefficiencies, their value becomes clear when used with the right expectations and validation strategies.
Recently, there’s been a surge of enthusiasm around the idea that anyone can build an application quickly using AI coding agents. I played a bit with base44 and the results are impressive! Curious to test this narrative myself without intermediaries, I decided to explore the capabilities of these tools in a more “personal” setting.
I’m now on vacation, so I’ve taken on a side project: recreating an AI-driven backgammon player, something I did many years ago (yes, this is what I like to do on vacations, outside of going to the beach, drinking margaritas, doing treks, and other stuff). This time, however, I’m relying entirely on AI agents for the implementation. I’m calling this approach “AI Coding”, a term I prefer over the more ambiguous “vibe coding.” In this experiment, I’m responsible for product specification and high-level architecture, while the AI selects the programming languages and frameworks. I can read most popular languages, so if the agent does really dumb stuff (and it does sometimes), I can fix them. Otherwise, I let it proceed autonomously.
Over the past two days, I’ve spent about two hours per day on this project. One of the first things I noticed is how different the time-on-task feels compared to traditional coding. When I write code myself, I usually have a clear sense of the scope and duration of each task. With AI Coding, this predictability disappears. Tasks that seem trivial, like changing the background color of a rectangle, can take multiple iterations. At the same time, more complex tasks, such as animating a pair of dice and triggering a roll on user interaction, are completed correctly in a single attempt. Still trying to understand why.
Another key insight is the importance of writing good prompts (as many people have written already). Instructions must be very specific, which sometimes make them very long and tedious. But vague or overly complex prompts often lead to partial or wrong implementations, misinterpretations, or complete failures. This unpredictability can be frustrating.
I’m using both GitHub Copilot and Cline, switching between them when one or the other gets stuck and can’t complete the task. It’s very important to configure these tools correctly. However, maintaining consistent global settings across the tools is not straightforward. I’m still refining my workflow, and the learning process has been enjoyable.
At times, AI coding feels like doomscrolling. You write a prompt, the agent responds with a working solution—great! Then the next prompt breaks everything. You revise it, and things get worse. You revert, try again, and slowly inch toward a solution. What should have taken two minutes ends up consuming half an hour.
Am I using the wrong tools? I don’t think so. Many developers report success with the same platforms. Am I using them incorrectly? Possibly. I’ve already seen noticeable improvements in my workflow after just a few hours of focused use. Time will tell – I’ve been doing this for a relatively short time, and I think I’m getting the hang of it. But there is still a lot to learn.
Final thoughts: AI-assisted software development is not magic, but it is powerful. It requires a shift in mindset and a willingness to experiment. With the right approach, I’m sure it can become a great tool, as others have demonstrated.
The journey continues…