How it works

v0.10.7

A tour of what Leaven, the Claude starter pack, does under the hood, for when you want to know why Claude does what it does, or to bend it to something new. No coding needed; every file named here is plain text you can open and read.

Two ways to do anything here

Every command in these docs can be typed by hand, or asked for in plain words: "make a folder for my Spanish course", "unpack the export in my Downloads". Claude runs the same command and shows it to you, which is a good way to learn the commands. Two things stay with you:

  • A command that asks for your password: anything starting with sudo, and the installer on Windows (WSL) and Linux whenever it installs system packages (the zsh, tools, and mise parts, and Node and poppler with job search ticked). On a Mac it asks only when it switches your login shell to zsh. Claude can't type a password, so run those yourself.
  • Starting Claude in a folder. Claude works in the folder it was started in, so to switch, type /exit, cd to the new folder, and run claude again. Claude can tell you the cd command.

What Claude Code is

Claude Code is Claude with hands. It runs in a folder on your computer, and besides talking it can read files, write files, and run commands there. Before it does something that changes your machine, it asks; you can answer yes once, or yes for that kind of action from then on.

Three ideas explain most of its behaviour:

  • The folder you start it in matters. cd ~/memory && claude gives it the vault; cd ~/job-search && claude, if you added the optional job-search workspace, gives it that workspace and its own rules.
  • A conversation has a limited working memory, the context window. Everything said and every file read takes up room in it. If you ticked the statusline part, the bar in the status line shows how full it is. When it fills up, Claude summarises the older part to make room, and detail is lost. Start a fresh conversation (/clear) when you switch to an unrelated task.
  • Nothing carries over between conversations unless it is written to a file. Claude starts every session knowing nothing about you. The vault exists to fix that.

What Claude reads when a session starts

Before you type anything, Claude has already been handed:

  • ~/.claude/CLAUDE.md: standing instructions for every session. The installer wrote its Memory section, which tells Claude to use the vault and to back up your work. You can add your own instructions below it ("I live in Denver", "keep answers short").
  • The folder's own CLAUDE.md, if it has one. In the optional job-search workspace, ~/job-search/CLAUDE.md points at AGENTS.md, and the rules in ~/job-search/.claude/rules/ say how to write without sounding like AI.
  • The vault's Bearing.md, pasted in by a hook (more on hooks below), so Claude knows what you are working on now.
  • A one-line description of every skill it has.

Ask "what instructions did you start this session with?" to see it for yourself.