Claude Code. What it does beyond a chat and the one mode that stops asking permission

A developer types one sentence into the terminal. "The login tests are failing, find out why and fix it." Claude Code opens the files, runs the tests, finds the bug in a completely different module, fixes two files and runs the tests again. Before each change it asks for permission.

In a chat the same developer would paste the code in piece by piece and copy the fixes back by hand. That's the difference between Claude Code and a chat.

What Claude Code is

Claude Code is Anthropic's tool for working with code. It reads the whole project, edits files, runs commands and works with git, so it can commit, create branches and open pull requests on its own. It runs in the terminal, in VS Code and JetBrains editors, in the Claude desktop app and in the browser. With the --add-dir flag you can give it access to a second folder, for example a separate backend repository.

You need a Pro, Max, Team or Enterprise plan, or an Anthropic Console account. The usage limit is shared with the chat, so a long Claude Code session uses the same limit as your conversations in the browser.

A business owner usually won't need it. But if you have a developer, it helps to know what this tool can do with your code and how to set it up. We wrote more generally about the difference between a chat and an agent in our article on AI agents.

Permission modes decide what Claude Code asks about

This is the most important setting. It decides what Claude Code does on its own and what it asks about first. In the terminal you switch modes with Shift+Tab, and the current one shows in the status bar. The documentation describes six modes:

• Manual, stored in settings files as default. Claude Code asks before most actions that edit files, run shell commands or reach the network. It only reads without asking.
• acceptEdits. It edits files and runs common file operations, like creating a folder or copying a file, on its own. It asks about the rest.
• plan. It only explores the project and proposes a plan. More on that below.
• auto. It works without asking, and a second model checks its steps in the background. By default it blocks things like force pushes, downloading and running scripts from the internet, irreversibly deleting files that existed before the session, and production deploys. Anthropic notes that this mode reduces prompts but doesn't guarantee safety.
• dontAsk. It does only what was approved in advance and denies anything that would need a prompt. It's meant for scripts and CI.
• bypassPermissions. It turns off prompts and safety checks. Every command runs immediately.

The mode that asks nothing

The bypassPermissions mode can only be turned on at startup, with --dangerously-skip-permissions or --permission-mode bypassPermissions. The flag name already tells you it's dangerous. Anthropic says to use it only in an isolated environment, a container or a virtual machine without internet access, where Claude Code can't damage your system. This mode offers no protection against prompt injection, an instruction hidden in a file or web page that Claude reads.

Picture Claude reading the README of a third-party library where someone has left such an instruction. In Manual mode you see a permission prompt for a strange command. In bypassPermissions the command simply runs. Claude Code refuses to start in this mode as root or under sudo, and on the first start you have to confirm that you accept responsibility.

If there are too many prompts, the documentation points to auto mode, not to bypassPermissions.

Plan mode for larger changes

In plan mode Claude reads files, runs commands to explore the project and writes a plan. It doesn't touch your code until you approve the plan. You enter it with Shift+Tab or with the /plan command followed by the task.

/plan add a CSV export of orders to the admin panel

Once you approve, Claude starts editing. Ctrl+G opens the plan in your text editor, so you can fix it before the work starts. It's a good habit for changes that touch many files. You see the mistake in the plan rather than in finished code.

/context and /rewind keep the session under control

/context shows what fills the context window as a coloured grid and suggests what to trim. It also shows which instruction files, such as CLAUDE.md, were loaded. We explained how to write that file in our post on CLAUDE.md.

/rewind, or pressing Esc twice with an empty input, opens a list of the prompts you sent in this session. At the prompt you pick, you can restore the code, the conversation or both. You can also summarise the conversation from that point to free up context. Claude Code creates a checkpoint like this with every prompt.

There are two limits to keep in mind. Rewind only undoes changes made with the file editing tools. If Claude ran rm, mv or cp in the terminal, rewind won't undo it.

Anthropic also says plainly that checkpoints don't replace git. A commit before a bigger task is still a good habit.

If your developer is starting out with Claude Code, have them spend the first weeks in Manual or plan mode and commit before every bigger task. Keep bypassPermissions for a container where there's nothing to lose.