Tools · Workflow

Claude Code, from zero to your first shipped change

Claude Code is an agentic coding tool. You run it inside one of your project folders, and instead of writing every line yourself, you describe what you want in plain language. It reads your actual files, plans the work, proposes edits and terminal commands, and once you approve, it writes the files, runs the commands, and makes the git commits. It is not autocomplete. It is a collaborator that operates your repo.

How the loop works

Every session is the same short cycle. The part worth internalizing is step three: nothing irreversible happens without your approval, so you stay in control while it does the tedious parts.

Install it (pick one)

You are on a Mac, so the native installer is the easiest path, it needs no other software. The other two options exist if you prefer them.

First run and sign-in

Open Terminal, move into a project, and start it:

cd ~/Developer/shawon-ch-learn
claude

The first time, it walks you through a one-time sign-in with your Anthropic account, the same Claude subscription you already use. After that, running claude in any project folder drops you into a session scoped to that repo.

The approval gate, and your safety net

When you ask for something, Claude Code shows you the exact file changes and commands before doing them, and waits for your yes. Review the diffs the first few times so you build trust in it. Your real safety net is git: because every change is a commit, anything it does can be reviewed and undone. Work in small steps and commit often, and you can never paint yourself into a corner.

CLAUDE.md is the memory

You already have a CLAUDE.md at the root of each repo. That file is how Claude Code knows your conventions without being re-told every session, your stack, your structure, your rules. It reads it automatically on startup. There is a hierarchy: a personal one at ~/.claude/CLAUDE.md, a project one at ./CLAUDE.md, and even per-folder ones, merged from general to specific. If you start a repo without one, the command /init generates a draft by inspecting your code.

A handful of commands worth knowing

Getting good output fast

Three habits do most of the work. Give it one task at a time rather than a paragraph of mixed requests. Keep changes small and committed, so each step is easy to review and revert. And lean on CLAUDE.md, every rule you put there is a rule you never have to repeat. For anything large, ask it to plan first and show you the plan before it touches files.

Work one, then finish one

Worked example. In your learn repo, you open a session and type:

build and publish a guide on the
derivative as a rate of change

Claude Code reads CLAUDE.md, copies TEMPLATE.html to math-for-ai/derivative/index.html, fills in the hook, figure, worked example, and recall check, adds the card to its subject hub, and commits, pausing for your approval. You push, Cloudflare deploys, and it is live at learn.shawon.ch/math-for-ai/derivative/.

Your turn. What would you type to add a dark-mode toggle to this guide's styling? (Something like: "add a dark-mode toggle to the guide template, respecting the existing teal accent." One sentence of intent, and you review the diff.)

Why this is the engine, not a toy

This loop is the thing that makes everything else you built sustainable. Your site, your learn hub, your future apps, all of them publish by instruction now. The friction that kills most personal projects, the dozen manual steps between an idea and a live change, is exactly what Claude Code removes. Low friction is your consistency mechanism, and this is the tool that delivers it.

Source: the official Claude Code documentation at docs.claude.com/en/docs/claude-code. Commands and install methods change, so check there if something looks off.