Been awhile trying to figure out this AI / LLM thing. If I step back, and see what I actually do, it’s 2 things:
- sounding board
- code review
Maybe this is just my current role, not sure, but even using TyDD & TDD, I’ve found using various models, they just don’t code the way I would have coded.
Even investing in an ever growing context file, they either get the types wrong, or the implementation wrong, or both “work”, but I just don’t like their code, or they hallucinate and I run out of patience. I turn the code hints off.
For sounding board, I find it valuable because I’m familiar with the subject matter, but sometimes appreciate someone to talk to about advanced topics. Examples are creating Type Driven Development tutorials in an OOP centric, bad type system, Test Driven Development community.
Sonnet is ok, but I’ve enjoyed Opus 4.x a lot as it’s much more verbose, detailed, and goes over the nuance.
Code review, however, has been surprising. I was disappointed I spent a lot of time getting better coding with LLM’s and getting a large context file… only to eventually conclude “this isn’t for me” a few times I’ve tried.
However, those context files work wonders in code review. They spot things I miss, bring up things I may have brushed off or thought not a big deal, and reinforce things I’m looking for. They’re super fast, too.
So while I’ve given up (again) using AI / LLM’s to code, using them as a rubber duck / sounding board has helped organize my thoughts & training a lot, and I find them super helpful in code reviews (yes, I hate PR’s and would prefer trunk, but that won’t change for awhile).
Leave a Reply