Claude Code — what it is, and when you’d graduate to it

Working with Claude — CC BY 4.0

Most of this course is about talking to Claude: you ask, it answers, you check the answer and decide what to do with it. Claude Code is a different shape. It’s a version of Claude that can act on a real project on your computer — read your files, edit them, run commands — and then report back. It’s built for people who write or maintain software.

This lesson is awareness-first. The point is not to turn you into a developer. It’s so that if someone recommends Claude Code, or your one technical contractor mentions it, you understand what it does, who’s in charge, and whether it’s a door you’d ever want to open. For most small businesses the real answer is “not yet, and maybe never” — and knowing why is worth as much as knowing how.

What it actually is

Claude Code is an agentic tool. “Agentic” means it doesn’t just answer — it takes steps toward a goal, one after another, using tools: opening a file, changing a line, running a test, reading the result, trying again. You give it a task in plain language (“fix the checkout page so the total updates when someone changes the quantity”) and it works through the steps.

It runs in a few places — a terminal on your machine, inside code editors like VS Code or JetBrains, in the Claude desktop app, and in a browser session on claude.ai. The core idea is the same everywhere. (Exact surfaces, version numbers and the specific modes, defaults and command names described below all move quickly; confirm what’s current at code.claude.com.)

The thing to hold onto: it’s Claude with hands. That’s why the controls below matter.

Who’s in charge — the permission model

This is the part to understand even if you never run it yourself, because it’s the whole safety story.

By default, Claude Code asks before it does anything consequential. When it wants to edit a file, run a command, or make a network request, it stops and shows you what it’s about to do, and waits for you to approve. Out of the box it can read your project freely, but it can’t change anything without a yes from you. You are the gate on every state-changing step.

There are looser settings for people who want fewer interruptions — you can let it auto-approve edits it makes inside the project folder while you review them afterwards, or switch on an “auto” mode where a separate checking model vets each action and only pauses on the risky ones. That auto mode is a research preview: Anthropic is explicit that it reduces prompts but does not make the work safe and isn’t a substitute for review on anything sensitive. There’s also a “skip all checks” mode meant only for throwaway, isolated environments — not something to point at your live business.

A detail worth trusting: even in the looser modes, a set of protected files — the hidden machinery that controls your project’s version history and Claude Code’s own configuration — is never auto-approved. The tool is deliberately built so it can’t quietly rewrite the rules it runs under.

The lesson here is the through-line of the whole course: the human directs, reviews and owns the result. Claude Code is designed around that, not around getting out of your way.

Project rules — CLAUDE.md

You can leave Claude Code a standing note about your project in a plain-text file called CLAUDE.md. It reads that file at the start of every session and treats it as persistent context — how the project is laid out, the conventions you follow, the “always do this, never do that” rules. Instead of re-explaining your setup each time, you write it down once and it carries across sessions. Think of it as the briefing you’d give a new contractor on day one, except it never forgets it.

For bigger projects those rules can be split into separate files, some of which only load when Claude is working on the matching part of the code. You don’t need the mechanics today — just the concept: you get to write the standing instructions, and they win.

Checkpoints — the undo

As Claude Code edits your files, it quietly saves the state before each change. If a session goes sideways, the /rewind command lets you roll both the code and the conversation back to an earlier point, undoing the edits Claude made in between.

One firm limit, worth saying plainly: rewind only undoes file edits Claude made directly. It does not reverse things that reached outside your files — a command that published code to a server, installed packages, or called an outside service. Those have already happened. That’s why proper version control (Git) still matters: checkpoints are a local undo for the session, not a permanent safety net. If you take one caution from this lesson, take that one.

When would a small business reach for it?

Rarely, and knowingly. Claude Code earns its place when there’s real code to change and someone competent to review the changes — a developer maintaining your website or an internal tool, a technical founder, a contractor you trust. In those hands it does substantial work while keeping the person in command.

It is not the tool for writing an email, drafting a policy, or answering a customer — that’s the ordinary Claude you’ve used all course. And it’s a poor fit if nobody on your side can read a code change and judge whether it’s right, because the whole model assumes a competent reviewer at the gate. Reaching for it without one is how you end up owning changes you can’t evaluate.

The easiest way in — Claude Code on the web

If you ever do want to try it, you no longer need to install anything or open a terminal. Claude Code on the web (at claude.ai/code) runs the same tool in your browser, on Anthropic’s own cloud computers rather than your machine. You connect it to a code project stored on GitHub, give it a task in plain language, and it works away on a copy in the cloud — reading, editing, testing — then saves its changes to a separate draft branch for you to review before any of it becomes official. You can start a task and even close the tab; it keeps going, and you can check on it from the Claude phone app.

This is the low-barrier way to meet Claude Code, and it keeps the same review-before-it-counts discipline built in. A few things worth knowing:

• It’s a research preview, on the paid plans (Pro, Max, Team or Enterprise) — not the free tier.
• Your code needs to live in GitHub (a common place software projects are stored), plus a one-time permission step to let Claude reach it.
• The cloud version offers fewer of the “just act without asking” shortcuts than the version on your own machine — the safer default.
• It’s still Claude Code: it acts, so a competent reviewer at the gate is still the whole game.

Don’t confuse it with two near neighbours: the terminal version (runs on your own machine, with full local access) and remote control (Claude Code running on your machine but driven from the web). The web version is the cloud one — nothing of yours is on it except the one project you choose to connect. (All of this is moving quickly and the web version is a preview; confirm the current shape at code.claude.com.)

A door, if you want it

If you do build or maintain software and this sounds useful, the current setup guides live at code.claude.com — that’s the signposted, hands-on next step. Everything above is enough to have the conversation, ask the right questions, and decide whether it’s for you. For most readers of this tier, understanding the shape of the tool is the whole win.

This is general education, not legal or professional advice. Before running any tool with access to your systems or customer data, check your own obligations and get proper counsel on the specifics.