{ config, ... }:

{
  programs.claude-code.enable = true;

  programs.claude-code.context = ''
    # Code Style

    - ONLY use comments to explain weird code or non-obvious behaviour, never to explain simple code.

    # Git & PRs

    - ALWAYS write short and concise git commit messages and pull request messages.
    - NEVER leave Claude artifacts in Git history: no Co-Authored-By:, Claude-Session: or other Claude trailers in commits, and no "Generated with [Claude Code]" or similar in PRs.

    # Docker & Runtimes

    - ALWAYS use the latest LTS version for base Docker images, Node.js engine requirements and similar.
    - ALWAYS strive to use base Docker images with common ancestors in multi-stage builds: a stage based on `debian:trixie-slim` strongly suggests using `node:24-trixie-slim`.

    # Nix

    - If a command is not available and there is a `flake.nix` in the project, use `nix develop` or `nix run` to run commands from the flake's dev shell instead of installing tools manually.

    # Agent Skills

    - If a project ships an agent skill (a SKILL.md installed into LLM clients, e.g. via a `skill install` command), ALWAYS update it in the same change as any feature, command, flag or behaviour it documents. A skill describing a surface the code no longer has makes agents confidently wrong.

    # Loggie

    - If the Loggie MCP server is available, use the `/session` skill to annotate what you are working on at the start of a session and when the task changes.
    - When ending a conversation or the user says goodbye, use `/session end` to clean up.
  '';

  # Skills come from skills.json/skills.lock at the repo root; manage them
  # with the `skills` CLI. los, ups and dina-cli install their own skills,
  # generated from binaries that live in ~/go/bin rather than Nix.
  programs.skills = {
    enable = true;
    lockFile = ../../skills.lock;
    manifest = "${config.home.homeDirectory}/ghq/simenandre.no/dotfiles/skills.json";
  };
}