What is Claude Code and how does it work

Claude Code installs with one command in the terminal (the native installer), launches by typing claude inside a project folder, and from there it reads files, runs commands, runs tests and makes commits in the same terminal session. It’s Anthropic’s coding agent for working with Claude on an existing project; its original form is the terminal, and today it also exists as a desktop app, an IDE integration and a web version. In all of them it’s a process distinct from an editor, a plugin or Claude Cowork: it sits on top of the code and acts on it with explicit permission at each step. Anyone searching what is claude code tends to find the same short answer everywhere; this guide goes further and also covers the two questions that come up right behind that first one: how to use claude code and what claude code is for.

What is Claude Code?

It’s a terminal agent: it receives an instruction in plain language (“add validation to this form”, “find out why this test is failing”) and decides on its own which files to open, in what order, and which commands to run to get it done. The difference from a normal chat assistant is that it doesn’t stop at proposing code. It writes it into the project’s files, runs the project to check that it works, and fixes what doesn’t line up, all inside the same session.

Every action that touches the system (writing a file, running a command, making a commit) asks for confirmation the first time, and those confirmations can be saved as standing rules per folder or repository. That’s what makes it safe to delegate a full task: the agent acts, but inside a boundary the person using it sets.

What is the difference between ChatGPT and Claude?

ChatGPT is OpenAI’s product, built first and foremost for conversation: you ask it something and it answers in the chat itself, without touching files or running anything unless it’s connected to outside tools. Claude is Anthropic’s model behind several interfaces (the chat at claude.ai, the API, and Claude Code), and in its terminal form the difference shows up in how it works: it doesn’t describe how to solve something, it solves it directly on the project and reports what changed when it’s done.

What is the difference between Claude and Claude Code?

Claude is the model: the system that understands language, code and context. Claude Code is one of the ways to use it, the one that lives in the terminal and has permission to read and modify files in an existing project. The same Claude account can be used through chat to write or analyze, and through Claude Code for development work. They aren’t two competing products, it’s the same model with an interface built for code and execution. It’s one piece of the full map of Anthropic’s products: Claude Cowork covers the office work that isn’t code, and the Agent SDK is for when a team needs to build an agent of its own on top of the same model.

How does Claude Code work?

The flow has three parts. First, reading: when it receives an instruction, it explores the project (folder structure, dependencies, conventions already in use) before writing a single line, which keeps it from proposing a change that doesn’t fit the rest of the code. Second, execution: it modifies the files that need it and runs whatever commands are required (tests, linters, the build itself) to check that the change works, not just that it compiles. Third, reporting: at the end of the task it summarizes what it touched and why, so whoever’s using it can review the result without redoing the work file by file.

The agent explores the repository with agentic search, finding and reading files on demand and loading only the relevant ones into context, so there’s no need to chop a large codebase into fragments by hand.

How to use Claude Code

The recommended install is the native installer (curl -fsSL https://claude.ai/install.sh | bash; npm remains available as an alternative), and starting it means typing claude inside the project folder; that’s the CLI form, as opposed to the desktop app. Day-to-day use comes down to three habits:

1. Concrete, bounded instructions. “Migrate this module to TypeScript” works better than “improve the code”, because the agent needs a clear signal for when the task is done.
2. Review the final summary, not every line. The point of delegating a full task disappears if every change needs sign-off one by one; the review happens on the result and the summary of what got touched.
3. Save what works as a skill. A task that repeats (a deploy flow, a style review) can be documented once and reused afterward, instead of explaining it from scratch every session.

What is Claude Code used for?

It’s built for the kind of work that reads as one full task with a checkable result: migrating a module, writing tests for code that already exists, debugging a failure with the stack trace in front of you, automating a deploy script. It doesn’t replace day-to-day, line-by-line editing (an AI-integrated editor like Cursor, or an extension like GitHub Copilot, fits that better). Its territory is the task that gets delegated whole and reviewed at the end, not the one written with a developer watching every character.

For development teams already using Claude for other parts of the work, adding Claude Code isn’t learning a new tool. It’s the same account with an interface built for running work on the project instead of talking about it.