I run claude code with 'dangerously skip permissions' flag enabled, and just tell it what I want. I hardly even interact with the OS anymore. I literally just tell it "install X, fix problem Y".
I do recommend checking methods against another provider. Occasionally Claude will suggest X, and Sol is like WTF are you insane? Clear often and update the prompt with your latest findings, because there are so many opportunities for the agent to find red herrings that they'll fixate on.
As with all LLM work, if you know some sysadmin, the process goes a lot smoother because you can say specifically "Write me a systemd unit file that does X" instead of a vague "I want to automate this, how do I do it" and not have a gut feel for which approaches are jank.