For us it’s been auto-generating tests - we focus efforts on having the LLM write 1 test, manually verifying it. Then use this as context and tell the llm to extend to all space groups and crystal systems.

So we get code coverage without all the effort, it works well for well defined problems that can be verified with test.