My philosophy with these LLMs is "maximal read access" and "minimal write access"
When I can implement this patters, it seems to work well. Of course the data backing these LLMS is massive, so it helps to give it as much data about my systems as I can.
Then of course these LLMS are dumb and make lots of mistakes or are over agressive, so I restrict their write ability as much as I can.
commit ae25a53