LLM side quests
While implementing a feature, the model reads a lot of code it does not need and spots problems along the way. Then it forgets them. One line in CLAUDE.md makes it file those as GitHub issues instead.
About
Hey! I'm Ernest, and through this blog I'm sharing a small piece of myself with you. I hope it brings some value to your own thinking on the topics discussed here.
A few words about myself. I spent my childhood in a small city in central Poland. Later I bounced around the country, first for my studies and most recently to settle by the Baltic Sea. I live here with my lovely wife, two dogs and one fluffy ginger cat. I graduated in Applied Mathematics, which had been my main interest throughout my school years. After that I wanted to break into the IT world. I started in SQL support roles, moved through various Scala engineering positions, jumped into the Rust ecosystem, and tried my hand at leadership. Quite a lot for my six years of experience! I've had a huge opportunity to challenge my knowledge and skills in both technical and people-related areas. A great experience. And I want more!
Why am I writing this blog? I'm someone who grows in constantly moving, changing environments. I like the speed :) But it comes at a cost: it's sometimes harder for me to gather my thoughts in a well-organised way. So the blog is my way of practising exactly that.
I also want to see whether my ideas can take root and help others on their journeys, mostly in the IT industry but not only.
I hope you'll find something inspiring in this glimpse into my brain. Enjoy!
Posts
Domain-Driven Agents
LLMs look brilliant in a greenfield repo and mediocre in a legacy one. The difference is the codebase. The fix is to invest in it directly: split strategic from tactical work, keep the strategic half, and give the model a domain it can actually speak.
Rust Respects Your Time, Dude
The first hour with the borrow checker is punishment, and it buys every week after it. rustc doesn't just reject the code - it explains why and attaches the patch, so the twenty minutes a vague error would have cost you never get spent. In an AI workflow that same output stops being a courtesy and becomes a gate the model runs against itself.
Bounded Chaos
Shipping real features with a model you can't trust. You can't make an LLM deterministic - so stop trying. Build a deterministic containment vessel around it (spec, gates, a dependency graph it can't talk past) and keep pushing that boundary outward.