Headless mode answers once and leaves. The interactive console —
launched by just typing claude — is where you work with the agent:
a conversation that keeps context, shows you diffs before they land, and
asks permission as it goes.
Inside the console, lines starting with / are commands, not prompts:
| Command | What it does |
|---|---|
/init | Analyzes the project and writes CLAUDE.md — the project memory Claude reads every session |
/model | Switches models mid-conversation (opus, sonnet, haiku) |
/help | Lists everything else |
/exit | Leaves the console (Ctrl+C twice also works) |
Always launch claude from the project directory. The console works
relative to where you started it: that's where /init writes CLAUDE.md,
and the files it reads and edits are resolved from there.
/model is a cost/power dial: haiku for quick questions, sonnet for everyday
work, opus for the hard problems. In this lab those map to real gateway
models — switching is free to try and metered against your budget.
Next: create a project and let /init document it.