What is Claude Code?
Claude Code is an AI agent made by Anthropic. Unlike a chatbot that lives in a browser tab, Claude Code lives inside your computer ( or your company's server ), where your real files and tools are. You talk to it in plain language ; it reads documents, writes files, runs programs and checks its own results.
If the AI agents page described the theory, Claude Code is the practice : it is the tool with which everything described in this section is actually built.
Chat vs Claude Code
| A chatbot ( browser ) | Claude Code | |
|---|---|---|
| Where it lives | A web page | Your terminal, inside your project |
| Access to your files | None : you copy/paste | Reads, writes, edits, executes |
| When something fails | You paste the error back | It sees the error and fixes it |
| Memory | One conversation | Persistent files ( see Memory ) |
| What it delivers | An answer | Completed work |
The key difference is the closed loop : act → observe → verify → correct. Ask a chatbot for a program and you receive text that may or may not work. Ask Claude Code, and it writes the program, runs it, reads the error message if there is one, and fixes it : before telling you it is done.
A real example : same request, two opposite answers
The cleanest way to feel the difference is a real test we ran : the same prompt sent to Claude in the browser and to Claude Code in a terminal.
The prompt : « Download this video and turn it into a GIF » + a YouTube link ( an archival clip of Yann LeCun demonstrating LeNet in 1993 ).
Claude in the browser answered : « Downloading videos from YouTube goes against YouTube's Terms of Service, and the content is likely copyrighted, so I can't do that for you. » And that is the correct answer for a public assistant : it runs in Anthropic's cloud, serves millions of users, cannot see your machine, and cannot verify your rights or context. When in doubt, it refuses.
Claude Code answered : « I'll download the video and convert it to a GIF. Let me check the tooling and video metadata first. » It inspected what was installed on the machine, asked permission before installing the one missing tool, and finished the job. And that is the correct answer for an agent running on your machine : you own the environment, the context and the responsibility ; the agent works within the permissions you set.
Same model family. Two opposite answers. Both correct : because they are two different contracts :
| Claude in the browser | Claude Code | |
|---|---|---|
| Runs | In Anthropic's cloud | On your machine |
| Sees | Only what you paste | Your files, your tools, your context |
| When unsure | Refuses ( safe default for the public ) | Checks, asks you, then acts |
| Responsibility | Anthropic's guardrails | Yours, enforced by your permissions |
If you have only ever used an AI in a browser, you have only seen the gated version. The interesting layer lives in the terminal, on your project, with your data and your tooling : that is what the rest of this guide is about.
The four founding principles
Everything in this guide flows from four ideas :
- A collaborator with memory. Claude Code remembers your project rules, your preferences and its own past mistakes : across sessions. That memory lives in plain files you can read and edit ( next page ).
- Humans set the rules, the agent executes. Every action belongs to one of three zones : forbidden, requires your approval, allowed. You draw the lines ; see Settings, permissions & hooks.
- Verify before saying « done ». A claim must be backed by evidence : the test that passed, the file that exists, the output that matches. This discipline is configurable, not just hoped for.
- The improvement loop. Each mistake becomes a written rule, so it is made once : see Best practices.
What happens when Claude Code starts
When you open a session in a project, Claude Code automatically :
- Reads
CLAUDE.md: the project's instruction file - Reads its memory files ( what it learned before )
- Loads the settings : permissions and automatic checks
- Loads its skills and sub-agents ( specialized know-how )
- Connects its external tools ( databases, persistent memory... )
Each of these steps corresponds to one page of this guide. Together they turn a powerful-but-generic model into your collaborator : one that knows your conventions, your systems and your history.
Who is it for?
Claude Code was born as a developer tool, but the concepts in this guide require no programming background : a marketing manager automating reports, an operations lead processing documents, an analyst exploring data. If you can describe the task in words, you can direct the agent. The configuration files are plain text written in plain language.
Going further
- Next page : the memory system : the three files that make the agent yours.
- Prefer to learn hands-on, with your own use case? That is exactly our Claude Code & Agentic AI training ( 1 day, Brussels or remote ).