After months of coding with LLMs, I'm going back to using my brain
15 May 2025
TLDR: LLMs are okay at coding, but at scale they build jumbled messes. I’ve scaled back my use of AI when coding and gone back to using my brain and pen and paper.
A few months ago I needed to build a new infrastructure for my SaaS, as the current PHP+MySQL combo was not fit for purpose anymore. I was excited about the opportunity to make the most of all the new LLMs I'd been playing with, so I set aside my SWE hat and I started acting as a product manager, chatting with Claude about best practices, doing some research on my own and then coming up with a plan, after many back and forths. I ended up choosing Go+Clickhouse.
When it was time to start coding, I asked Claude to give me a big and complicated markdown file outlining my existing infrastructure, my desired new infrastructure, what I’m trying to achieve, why I'm doing it, etc.
So I put it all inside Cursor Notepads and I start prompting away. Cursor writes code, I build it and test it. I'm quite happy with how things are going, the codebase isn't the tidiest but things seems to work.
I'm aiming for speed more than code cleanliness - my SaaS business customers told me they need specific data and this new infrastructure is the only way I can deliver it. I have a few more potential customers who are waiting for me to tell them this is ready so they can purchase a plan. Every day I don’t have this ready, I’m literally losing money.
But then a few weeks go by, and the cracks start to show. I start getting frustrated. Every day I feel like I’m really close to the end product, then I find another issue that sets me back for days.
I justify it thinking that I’ve never used Go or Clickhouse before, so it makes sense that it takes me a bit longer than I’m used to to fix those issues. The problems continue though, and I’m getting more frustrated. Cursor is not helping as much anymore. I paste in error messages, and I get a fix in response but then something breaks somewhere else.
The more detailed the problem, the harder it is for the LLM to provide the actual solution. So I start actually looking at the code more closely, trying to understand it.
I’ve been a software engineer for 15 years, and I studied C++ and Java in school, so I do have a rough idea of what’s happening in these Go files. But I have no concept of Go or Clickhouse best practices.
I start educating myself a bit more about those. I read some documentation, articles, I watch a YouTube video about Clickhouse. I become more inquisitive with Claude, asking detailed questions and challenging its answers.
One morning, I decide to actually inspect closely what’s all this code that Cursor has been writing. It’s not like I was blindly prompting without looking at the end result, but I was optimizing for speed and I hadn’t actually sat down just to review the code. I was just building building building.
So I do a “coding review” session. And
the horror ensues.
https://albertofortin.com/writing/coding-with-ai
