My AI agent stack/workflow

NixOS

Best OS known to mankind.

• Declarative - AI agents can be trained more effectively on them and Claude Opus 4.8 nails it.
  • NOT your Dockerfile type dumb script
    • Hardware configurations
    • OS configurations
    • Firefox bookmarks
    • File manager keymaps
    • Frigate NVR
    • Coral Drivers
      • Enabled Option
      • Pre-requisites
    • Custom options
  • Declarative -> Version Controlled -> Reviewable -> Revertable
  • Not the biggest fan of nix language syntax though
• Reproducible
  • No more weeks of setting up a new system.
• Immutable

MicroVM

Safe space for agents to YOLO

• Why MicroVM and not docker/podman?
  • Docker sandbox is OK!
    • Agents does not know what the hell Docker Sandbox is
    • Docker sandbox couldn't make creating custom templates annoying enough
    • They suggest moving my user scope skills to the workspace itself
    • Why the hell do I have to mount a workspace to run a sandbox???? What kind of devcontainer BS is this?
• NixOS + microvm.nix - Life is beautiful
  • Reuse the same config you already have
  • Mount what you need not what they want

Headroom

Delay end of the world by 1 picosecond. Use less tokens

• Running Host -> MicroVM -> Claude -> Headroom Proxy Container -> Claude Server
  • This is ready and waiting for you. Auto starts with systemd. No commands/manual setup needed.

Neovim -> (Claude + Lazygit)

• Occasional code navigation alongside Claude
• Reference context
  • File references
  • Line range
• Review code changes locally

Usual workflow

• /grill-me or /grill-with-docs then /to-prd
• /task
• Manual PR review and merge

Settings

• Enable YOLO mode
• Tell claude to shut the hell up

Skills

• grill-me with minor changes
• to-prd
• git-commit
• task

Plugins

• LSP
• Context7
• playwright

Hooks

• Make sure everything becomes a PR at the end