Claude Code has to be actively steered, because while it knows some nixpkgs it surely doesn’t know it enough. E.g. it was absolutely incapable of fixing lldap settings after system upgrade from 25.05 to 25.11. It just prodded around blindly, producing meaningless configs instead learning how the module works.
NixOS docs work for me, but I tend to just go for the nixpkgs source instead. Manuals document options but not how those are actually plumbed through, nor what remains behind the scenes like all systemd unit settings). Claude can do this too, but it goes quite weird roundabout ways with a lot of weird `find /nix/store` and `nix eval`s to get to it, slow and token-hungry (and not always accurate).
This said, Claude is very helpful at checking logs and providing a picture of what’s going on - saves ton of time this way. Plus it can speed up iterating on changes after it’s fed enough knowledge (but don’t expect it to do things right, that’s still on you). It has breadth of it, but not the depth, and that shows at almost any non-trivial task.
I feel you on the nix store + nix eval death loop, though it gleans real info. If I weren't on the Claude Max plan I'd probably feel more of the pain. And context is now 1MM tokens which means you're not running out just as it's starting to piece things together, heh.
I’m going to experiment with skills next, or maybe make it build a few helper scripts for itself to quickly get some module source from nixpkgs matching flake.lock without having to think of it all. I’m positive about Claude for nix management, merely saying it’s not something that “just works” for now and reading nix code is still on the human part of the tandem.
This said, to be fair - when it gets the approach right, it excels. I was setting up Ente for photos backup and sharing, and it produced a nice overlay with custom patches for my needs from just “figure out why /shared-albums/ redirects wrong and fix”. Found the module, the package, pulled source, analyzed it, proposed a patch (settings weren’t enough), did it - I only had to test, and only because I haven’t provided it with a browser. Felt amazing.
I think the initial migration towards nixOS is the hardest point, since it requires learning a bunch of new things all at once in order to get the system into a usable state that matches your expectations and preferences. The key benefit of using an LLM is that it makes it really easy to get your system into a useful initial state, and then you can safely learn and experiment incrementally with a mix of tools.
When I started off I didn't understand everything, but at this point I feel I have a very good understanding of everything in my configuration file.
Unless you're brand new to Linux or computing, it's not a mystery what a given nix config change is ever doing.
You can probably guess what this does:
networking.firewall.allowedTCPPorts = [ 8080, 9000 ];
I am not brand new - and I don't know what the heck the config is doing.
That is why I rely on documentation.
The "code is self-explanatory" is always an attempt to not have useful documentation and try to rationalise that problem away.
You can read documentation on an as-needed basis or to your heart's content.
The point is that the majority of the day to day changes I make to my desktop environment aren't so critical that I need to do more than read an AI agent's proposed changes to my config and accept them when they look reasonable.
And I don't think looking up the exact config options to NixOS' networking system does anything to increase my knowledge of the OS. It's just a triviality.
Like what is ultimately the difference here for you vs a non-nix user who, as author says, is just dealing with some big ambiguous pile of state? It kind of takes away any upside to using nix, and probably just creates more friction for your AI than just running ubuntu/apt stuff.
The idea is you can keep configuration "in your head" such that you can reason and iterate and fully know what your system is like at any moment. If you actually don't care about that, you aren't getting anything out of it!
I have these packages installed and these firewall settings and these users with these permissions and this folder served over Samba and these hotkeys that do these things and these Obsidian vaults synced over SyncThing and these devices in my SyncThing network and Neovim installed with these plugins and ...
This is difference between me and a non-nix user, not whether we can rattle off the exact state of our live system from memory.
The non-nix user has to query live system state, if such query tools even exist for their question, and I get to read a config file. And I get to maintain my system config in git, and I get to deploy my config on all of my machines.