My least favorite topic in software engineering is, at any given time, the subject of the current hype cycle. At least LLMs are more useful than blockchain! Truly though, the last three years have really reworked how almost all of us write software.

That's both good and bad. For every instance of the LLMs saving me time or being genuinely helpful, I'm sent an incomprehensible 30-file PR that, in addition to probably not doing what the "author" thinks it's doing, is littered with the most annoying emoji-laden comments.

I think we're all generally agreed that using LLMs as a crutch isn't an acceptable way to go about engineering software. The tools have proven good at digging themselves into a hole but not good at the sort of long-term quality thought that software engineers are typically ignored for. That said, it's difficult to form concrete practices from general statements like that. I wanted to develop more concrete thoughts on what the line is between productive and unproductive LLM use.

So, my reading these last three months has largely been: going back over some of my favorite authors' thoughts on these LLMs. Pro tip: having a blogroll is a huge help when I'm looking into any topic! Here's what I've collected, in alphabetic order by author name:

• https://brandur.org/nanoglyphs/048-llms
• https://brandur.org/minimum-viable-unit
• https://benhoyt.com/writings/indecisive-ai-agent/
• https://benhoyt.com/writings/how-to-apply/
• https://www.ben-morris.com/software-is-becoming-easier-to-write-than-to-understand/
• https://www.ben-morris.com/are-we-using-ai-to-automate-the-wrong-things/
• https://www.ben-morris.com/why-building-ai-powered-agents-is-so-challenging-for-now/
• https://lemire.me/blog/2026/07/16/using-ai-to-build-your-own-software/
• https://blog.maartenballiauw.be/posts/2023-06-15-running-large-language-models-locally-your-own-chatgpt-like-ai-in-csharp/
• https://martin.kleppmann.com/2025/12/08/ai-formal-verification.html
• https://www.michaelnygard.com/blog/2026/05/ai-versus-microservices/
• https://www.michaelnygard.com/blog/2026/05/ai-versus-throughput/