cannedtunaMathieu Comandon Explains His Use of AI in Lutris Development
I was ready to agree with a pragmatic use of LLM’s until he mentioned that he is working in an agentic style and is scared to touch the code manually because it will confuse Claude. 🤦🏻♂️
I’m all for using the AI as a pair programmer but unleashing it on a gigantic Python codebase is a comically foolish move, IMO. Python is not safe for this method of working. In fact, I’d call anything short of Unison (which hashes the entire AST) or something tailor-made for LLM’s like aver a fool’s errand. In my experience, the LLM needs a strict set of guardrails (like a static type system or similar) to get feedback from lest it turn things to spaghetti.
because it will confuse Claude. 🤦🏻♂️
You’re kidding? Or did he really wrote that?
Do you still review and rewrite the code that Claude produces before it lands in Lutris?
Mathieu: I do review the code but I usually don’t rewrite something Claude is actively working on. If you don’t tell Claude about any changes you’ve made, there’s a big chance it will revert those and switch back to its original implementation! Instead, I will prompt Claude for the fixes I want made until I’m happy with the result.
Wait, that is completely different from what you said earlier.
“if you don’t tell Claude, it will override your change”
Agentic tools do this today.
imagine this, if you set a goal for a robot to build a 9th storeys building, later you interject mid way to build up to 15 storeys. When the robot resumes its task, of course the robot is going to destroy the higher floors to meet its original goal of building to 9th storeys.
what else should we expect here?
the right to do before hitting the resume button is to “communicate” this new intention again.
it is not about being afraid of things