Don't forget that LLMs are trained on human code. If they cannot understand what your code does then they cannot make changes to it, or at least - having them understand your codebase becomes expensive (more trips to Anthropic servers)