After trying a few different LLMs, I'm starting to think that letting them directly modify your project's code leads to non-deterministic output and knowledge that’s hard to reuse.
A better pattern might be using LLMs to build tools that then apply those changes in a deterministic way.
The knowledge stays in the tool instead of disappearing when the session ends.
Next thing I'll try to share these tools with others, or maybe letting an agent run them at scale.
