jak0

Yes absolutely! Please have a look at my nixos-config I really would love to get feedback from others

That's a real scenario and it's happened to me. The key is that NixOS keeps every previous generation — so when I discover program #12 is broken, I can nix store diff-closures to see exactly what changed between generations, or just boot into the last known-good generation from GRUB. In practice I do two things: I keep a quick smoke-test habit after rebuilds (open a couple of the less-used apps), and the git history tells me exactly which .nix change touched what. Since every rebuild is a commit, git bisect works on your OS the same way it works on code. It's not perfect — I've definitely had the "wait, when did this break?" moment. But the recovery path is always under a minute, which makes the cost of experimentation very low.

Ahahaha yes I totally agree - but my point is that nixos seems to be the right candidate to at least try to manage the entropy :) Or maybe it just gives you this impression.

I will definitely look into guix thanks!

NixOS + Flakes turns your entire operating system into a folder of .nix files. That folder is a git repo. Give Claude Code access to that repo and it can manage everything — packages, services, hardware config, shell, bootloader. I've been calling this ClaudeOS: an operating system entirely managed by Claude Code.

I'm not a developer. My background is data science and finance. Six months ago I couldn't tell you what a Nix derivation was.

I started because I wanted Claude Code to manage my system, not just my code. Tried it on Ubuntu first and it was a disaster — Claude would edit .bashrc and break my shell, install packages that conflicted with each other, no way to undo any of it cleanly. A friend mentioned NixOS, I installed it on my Framework laptop, and within a week I realized this is what AI-assisted system management should look like. Claude edits a .nix file, I rebuild, if it breaks I roll back in one command. No more "what did the AI just do to my system?"

It snowballed from there. ~470 commits later I'm running 7 machines off this config — my Framework 16 dev workstation, a ThinkPad, and a bunch of business laptops for people who've never opened a terminal. Two profiles: a "tech" profile with 350+ packages and a full AI toolchain (Claude Code, Cursor, local speech-to-text), and a "business" profile with ~40 curated packages for office use. Adding a new machine is three lines in flake.nix.

Some things I built along the way:

A script that spins up Claude Code in a sandboxed git worktree with bubblewrap + seccomp so it can work autonomously in the background. It runs in a tmux session and loops with fresh context up to 5 times. I use it for overnight refactoring.

Custom NixOS installer ISOs — I ship a USB stick to someone, they plug it in, and they get a working system with Claude Code pre-configured as their "sysadmin." They ask Claude to install software, Claude edits the config, they rebuild. I manage their machines remotely via git push.

CI/CD with BATS tests, ShellCheck, security scanning. A two-branch model where personal (my dev branch) auto-syncs to master via CI with path sanitization so nothing personal leaks to the public repo.

The core insight: NixOS is the only OS I've used where AI can't permanently break anything. Declarative config means Claude always knows the exact system state. Atomic upgrades mean every rebuild either succeeds completely or doesn't happen. If something goes wrong, I pick the previous generation from the boot menu. I've bricked my system maybe 15 times and recovered in under a minute every time.

What still sucks: the Nix learning curve is real even with AI. Claude writes non-idiomatic Nix all the time and I can't always tell. Flake lock updates break things in ways that take hours to debug. Error messages are famously terrible. And NixOS is not for everyone — it's a tradeoff between upfront complexity and long-term reliability.

Is anyone else doing something like this? Not just using AI to write code, but to manage and evolve their actual operating system?

Repo: https://github.com/jacopone/nixos-config

Can this be tweaked somehow to try to reproduce the experience of Aqua Voice? https://withaqua.com/

If you guys mostly have CS background you should definitely check this out : http://www-stat.stanford.edu/~tibs/ElemStatLearn/

It's Data Mining from a statistician point of view. You can download the entire book for free on the website and the graphs and computations are done with R.

This book is also one of Hal Varian's favourites: http://www.dataspora.com/blog/sexy-data-geeks/