Class 6 · 6.2

Getting started with the CLI: init, context & testing

Open a project, launch Claude, and meet the commands you'll use every day — /login, /init and /context — then let it run your whole test suite.

🎯 Goal: drive the Claude Code CLI⏱ Read: 8 min🧭 Type: hands-on

01Open a project and launch Claude

Open the project you built last week (our PM app) with File → Open. Before doing anything, it's worth committing your work so you have a clean baseline — then let Claude loose on it.

VS Code · terminal

git status
git add .
git commit -m "step 10 complete"
# start absolutely fresh, then launch:
exit
# open a new terminal (the + button), then:
claude

Up comes the Claude Code CLI — characters and lines printed in a very old-school way, like a previous decade of computing. The throwback feel is intentional: it's meant to feel like a raw, native interaction with the LLM. To talk to Claude, just type. To give it a command, start with a slash (/).

02`/login` and `/init`

Your first command is /login — it opens a web page to sign in with your Anthropic credentials, then drops you back logged in. (If you want free models instead, hang on until later.)

Claude uses CLAUDE.md, not agents.md

Claude Code has its own memory file, CLAUDE.md. You could simply rename your agents.md to CLAUDE.md — or run /init, which asks Claude to explore the project and organise itself.

Run /init and Claude shows one of its playful processing words, then searches, analyses and explores the directory structure — reading your existing agents.md along the way — before writing a CLAUDE.md. You'll see the diff in VS Code (even from the terminal, Claude can drive the editor), and then its famous edit prompt:

Press 1 — Yes

Accept this single edit.

Press 2 — Yes, and allow all this session

Shift+Tab is the shortcut. Your finger lives over this area when you use Claude Code.

Press 3 — No

Reject the edit. You can also press Tab to add a comment — "yes, but also think about…".

Don't let it write its own CLAUDE.md as a rule

As a general principle, write CLAUDE.md yourself — just as we did for agents.md. It's the most crucial thing for the human supervisor to get right, so invest your time there. Letting /init draft it and then improving it is fine; letting it own the file forever is not.

03`/context` — mind the window

A command you'll run constantly: /context gives a quick visual of how much of the context window is being used. Claude's window is 200,000 tokens — deliberately not huge. Anthropic wants a smart, powerful model with carefully managed context.

Memory + messages eat into a fixed 200k window; the buffer is kept free for compaction.

Memory holds things like your CLAUDE.md; messages are your interactions so far — even basic setup uses a fair amount. The buffer at the end is reserved room so Claude can maneuver.

Compaction

When the window fills, Claude compacts — compressing the conversation down to something smaller so it can keep going. The buffer is the wiggle room it needs to do that.

04Put it to work — run the tests

Now instruct Claude the way you learned last week. First, make sure it understands the project:

Please read plan.md in the docs folder to understand everything that's been built, and any supporting docs.

Then ask it to verify the project actually works:

Please run all tests to confirm everything is working — bringing the server up as needed and back down at the end.

You'll hit the familiar permission prompt (allow this docker build command?). Press 2 to say yes and not be asked again for docker builds. In one memorable moment, Docker wasn't running after a restart — and Claude knew how to launch Docker Desktop itself.

Opus is that good

Auto-launching Docker is the kind of thing earlier models couldn't reliably do. Claude ran every backend and frontend test, and reported them all passing — two harmless deprecation warnings aside. A quick /context afterward shows how fast messages fill the window, so keep a healthy eye on it.

✓ Key takeaways

Commit first, then launch Claude Code with claude in a fresh terminal — the retro CLI feel is intentional.
/login signs you in; /init explores the repo and drafts a CLAUDE.md — but own that file yourself.
The edit prompt: 1 yes · 2 yes & allow all · 3 no · Tab to add a comment.
/context visualises a fixed 200k-token window — memory, messages, free space and a compaction buffer.
Instruct Claude to read the docs and run the full test suite; a strong model can even start Docker for you.