The Programmer Identity Crisis

Ignoring LLMs for a second, some code I write is done in sort of full-craft full-diligence mode, where I am only committing something where I am very proud of it's structure and of every line of code. I know it inside and out, I have reasons for every decision, major or minor, and I don't know of any ways to make it better. Not only is the code excellent, I've also produced a person (me) who is an expert in that code.

Most code is not like that. Most code I want to get something done, and so I achieve something quite a bit below that bar. But some things I get to write in that way, and it is very rewarding to do so. It's my favorite code to write by a mile.

Back to LLMs - I find it is both easier than ever and harder than ever to write code in that mode. Easier than ever because, if I can actually get and stay in that mode psychologically, I can get the result I want faster, and the bar is higher. Even though I am able to write MUCH better code than an LLM is, I can write even better code with LLM assistance.

But it is harder than ever to get into that mode and stay in that mode. It is so easy to just skim LLM-generated code, and it looks good and it works. But it's bad code, maybe just a little bit at first, but it gets worse and worse the more you let through. Heck, sometimes it just starts out as not-excellent code, but every time you accept it without enough diligence the next output is worse. And by the time you notice it's often too late, you've slopped yourself, while also failing to produce an expert in the code that's been written.

Within the past 2 months, as I've started to use AI more, I've had this trajectory:

  1. only using AI for small things, very impressed by it
  2. giving AI bigger tasks and figuring out how to use it well for those bigger tasks
  3. full-agentic mode where AI just does its thing and I review the code at the end
  4. realising that I still need to think through all the code and that AI is not the shortcut I was hoping it to be (e.g. where I can give it a high-level plan and be reasonably satisfied with the final code)
  5. going back to giving AI small tasks

I've found AI is very useful for research, proof-of-concepts and throwaway code of "this works, but is completely unacceptable in production". It's work I tend to do anyway before I start tackling the final solution.

Big-picture coding is in my hands, but AI is good at filling in the logic for functions and helping out with other small things.