Theo Pavlakou

The best parts of using AI

Here are the top ways I've found AI helps me the most, which I believe aren't emphasised enough.

1. Understanding new complex code bases

One of the best ways I've noticed AI has helped me in my day-to-day role as a software engineer is that it's just so much easier to understand a new codebase now. Whether that's for onboarding to a new project or just because I need to navigate a new part of the stack to add a feature for myself, the friction associated with being able to be productive is just so much lower.

For instance, I've had situations where I'm working on some service, which after multiple hops communicates with some backend e.g. a logging service, and I want to trace all the potential paths that might make it from my service to the backend I care about e.g.

Without AI you can use tools like tracing, but that only shows you which path was executed, not all the possible ones. There are also some static analysis tools that are useful for certain languages, but they're not exhaustive and rarely work well across services implemented in different languages e.g. a Python service calling a C++ service.

With AI, I can typically get a good result at the level of detail I care about for the job at hand with a single prompt e.g. "Give me all possible codepaths that start at function/class A and end in service Z. Enumerate each path and list each of the functions that gets called in the chain." And that's usually good enough!

"But doesn't it hallucinate?"

Sometimes. But it still makes my job easier, because I can easily verify when it's done so and, just by providing a set of possible paths, it's limited my search space by orders of magnitude.

"Doesn't it miss some paths?"

Again: sometimes. But again: it's still far easier to do the job if 90% of it has been done and only needs to be verified.

The point isn't that I can now be lazy and trust the output 100%. It's just the fact that I can now focus 20% of my energy on verifying an output from an LLM (which is less daunting and provides equal value) and the remaining 80% can be used for parts that it might have missed and actually understanding the system at a high level.

2. Creating short courses on particular topics

Another area that has been a game changer for me has been in creating short educational courses for topics that I'm not that familiar with e.g. my own text-based Udemy type courses. I've used Claude Code for this extensively.

Something as simple as: "Create a 10-15 part course for me on LLMs, which includes some background on how they were discovered, how they're trained, fine-tuning methods and computation bottlenecks. Each chapter should build on previous chapters and at the end of each chapter there should be a short quiz." This can yield amazing results, and you can iterate on the style as you go through the course.

I've done a few of these until now and at the end of each chapter I ask it to save progress and to use my questions during the chapter to change its approach to presenting the content over time. For instance, if