Over the last year I have started to think of AI as something like a very fast junior developer. It can read and understand a codebase quickly and produce a lot of code in a short amount of time, but it needs clear instructions and you still have to review everything it writes. That comparison has ended up being the most accurate way I can describe what using AI in day to day development actually feels like.

Before AI became part of my workflow, most of my day looked like what you would expect for a software engineer. I spent most of my time writing code. The rest of the time was usually spent researching libraries, reading API documentation, or figuring out how to implement something. That research portion often took longer than the actual coding. Context switching between development and meetings was also a constant challenge, and that part has not really changed.

What has changed is how quickly I can move through implementation work. One of the first practical things I used AI for was porting a TypeScript library to PHP. I could have done it manually, but it would have taken a while. Instead I tried letting AI translate the code. To my surprise it worked fairly well. There were a few bugs where it generated incorrect PHP syntax or called the wrong methods, but the overall structure translated better than I expected. It even handled some of the differences in flow between TypeScript and PHP without much trouble. That was the first time AI stopped feeling like a novelty and started feeling like a useful development tool.

These days I use AI regularly to implement new features. My typical workflow is to describe the feature I want and the libraries involved, let the AI draft an implementation, and then iterate from there. It usually takes some back and forth to get things right, especially when the libraries involved are less common. Models tend to perform much better when there are plenty of examples available.

In practice I now write far less code by hand than I used to. Instead I spend more time reviewing and guiding the code AI generates. I still read every line it produces. AI is capable, but it is not careful in the way an experienced engineer is. It will call the wrong methods, apply patterns from unrelated projects, or misunderstand how a library is supposed to be used.

AI also struggles when a project is completely new. If there is no structure to follow, it tends to guess wrong. In those cases it is usually better for me to write the initial groundwork myself. Once the project has some shape and patterns in place, the AI can usually follow them after a little coaching.

Where AI helps the most is with repetitive work. Writing tests is a good example. I used to dread writing tests because they are repetitive and writing good tests can easily take longer than writing the implementation itself. AI handles that kind of work well and can generate a solid starting set of tests very quickly, which I then review and adjust. I still verify the tests carefully because AI has a habit of creating test doubles or mocking things in ways that avoid actually testing the real code, especially when things get complex.

AI has also changed how I approach learning new APIs and libraries. I tend to learn better from examples than from reading documentation. AI is very good at reading documentation and turning it into working examples, which makes it easier for me to start with a working snippet and adapt it from there.

Another task AI has made much easier is porting libraries from other languages into PHP. I have done this several times now. It is something I could have done before, but it would have been very time consuming. AI makes that kind of translation much faster.

The biggest practical change is speed. I can get a lot more code written in a day than I could before. Features move faster, and the code often ends up better documented and better tested because AI generates those pieces alongside the implementation. Because of that speed it has also changed how I estimate work. Tasks that used to take much longer now move fairly quickly, and since I tend to take on as much work as I can fit into a day, faster implementation usually just means I attempt more things.

AI is not perfect. It makes mistakes, it sometimes applies the wrong patterns, and it is not great at starting projects from scratch. But for many everyday development tasks it is genuinely useful. I could probably go back to writing software without AI, but right now I do not really want to. Using AI in development has turned out to be fun in a way I did not expect, and for the moment it has become a normal part of how I write software.