CLAUDE.md is basically a user manual that you hand to Claude.
It says how your project works, which rules apply – and what Claude must never do.
Exactly. Claude reads the file automatically as soon as you start inside the project.
There are even three levels: globally in ~/.claude/, in the project folder and per subfolder. /init writes you a first draft.
A skill teaches Claude how a particular job is done properly.
Instead of explaining your whole procedure every time, you turn it into a reusable skill.
Say you want the same analysis report every week.
Your instructions, your procedure, your examples – that becomes a skill you simply call.
Pretty much. A skill is a folder with a SKILL.md in it, plus scripts, templates and reference files.
Claude first reads only the name and description – the rest only when it really needs it. You can also call it yourself with /skill-name.
Not quite. A skill is one ability.
A plugin is more of a bundle: skills + hooks + subagents + MCP servers + custom commands in one go. You install it from a marketplace with /plugin.
MCP is basically how Claude connects to things outside of Claude.
One standard plug for other people’s programs and services.
Claude connects on its own, fetches the information – and may act there too, if you allow it.
/mcp shows you what is currently connected.
Hooks are different because Claude doesn’t decide whether they happen.
They fire automatically as soon as a certain event occurs. Set up in settings.json or through /hooks.
A hook can run your formatter after every file change – or stop Claude before a dangerous command runs.
Typical events: PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Stop.
They are independent workers Claude can hand clearly defined jobs to – each with its own context window and its own model.
One researches, one reviews code, one tests. Then they report back to the main Claude. You create them as files under .claude/agents/ – or you let Claude do it for you.
The slash is Claude Code’s shortcut menu.
There are built-in ones like /clear and /compact – and skills bring their own commands that you can call yourself.
On the paid plans Claude Code now starts in auto mode: it works straight through, and instead of you a second model reviews every action.
You can switch at any time with Shift+Tab: manual goes back to asking before every change, plan only looks and puts a proposal in front of you.
Four characters save most of the typing – and two keys are the emergency exit.
Remember Esc above all: you don’t have to wait until Claude is done. Stop it, put it right, carry on.
That is Claude’s working memory for the conversation you are having.
Your input, Claude’s answers, files it has read and tool results all take up room in it.
… then usually so much has piled up that it no longer all fits into the working memory at once.
Claude Code then summarises automatically. /context shows you what is taking up the space, /clear gives you a clean start.
No – Claude Code sets checkpoints automatically before it changes files.
Esc twice or /rewind, and you roll back: just the conversation, just the files, or both.
Yes – that is what memory is for: small note files Claude writes itself and reads again the next time it starts.
Your preferences, where a project stands, decisions that keep coming back. Unlike CLAUDE.md, Claude maintains this itself.
Not for a long time now. Claude Code comes as a terminal command, as a desktop app, in the browser at claude.ai/code and as an extension for VS Code and JetBrains.
On top of that, agents that keep working in the background or in the cloud while you do something else.
Yes. The Claude desktop app has three tabs: Chat is the conversation, Cowork does office work, Code is Claude Code with a user interface.
It is available for Mac and Windows, and as a beta for Linux.
Several sessions side by side – each in its own Git worktree. That way they don’t get in each other’s way, even when they work on the same spot.
Plus editor, terminal, preview and the review of the changes in the same window. You can even kick off a session from your phone.
That is what /batch is for: Claude thinks the rebuild through once, then hands it out to many agents that start at the same time – each in its own working copy.
At the end each one puts up its own set of changes. You see them side by side and accept them one at a time.
The same machinery as Claude Code – only in the ordinary Claude app instead of the terminal, with nothing to set up.
You share folders, describe the goal, and Claude gets going: spreadsheets with real formulas, presentations, finished documents.
Turn a folder full of receipt photos into an expense report. Write a report from a pile of notes. Tidy up a filing system. If need be it will drive Chrome for that – clicking, typing, filling in forms.
How much it may decide on its own is up to you.
The question is always how far Claude may go into your things. Cowork gets individual folders, runs sealed off and needs no setting up.
Claude Code gets the whole machine. Only here can you teach it tools of your own – and only here does it keep going when you are away.
Then you create a routine: task, project and accounts saved once – after that it runs without you.
And it does so in the cloud. Your machine may be switched off. You create it with /schedule, in the app or in the browser.
Three things, alone or combined: a schedule, a call from outside, or an event on GitHub.
The shortest interval is one hour. One-off appointments work too – /schedule in two weeks … The whole thing is still in preview.
Then use /loop. It repeats a task inside the open session – on your machine, with your files.
Leave out the interval and Claude picks it itself: short while something is happening, longer when things are quiet. Esc ends the loop.
The difference is where it runs. The loop needs your session open and your machine switched on – in return it sees your files.
The routine runs in the cloud, without you. Rule of thumb: loop for this afternoon, routine for every Monday.
Then start with /design. Claude lays several drafts out as boards on a canvas – click them, move them, edit the text right there.
Once the direction is right, you hand it back to Claude Code, and that is where it gets built. Still an early preview.
As of August 2026: Fable 5 is the strongest and built for long runs, Opus 5 takes the heavy lifting, Sonnet 5 the everyday work, Haiku 4.5 the fast and cheap jobs.
Switch with /model. With /fast Opus answers more quickly – it is not swapped for a smaller model in the process.
Yes, that is the second dial next to the model: /effort. First pick the model by how hard the task is, then the effort by how thorough you want it.
The default is high. Going down means faster and cheaper, going up means more thinking, more tool calls, longer runs.
State the goal, not the individual steps – Claude finds the way itself. And say how you can tell when it is done.
And give the context: who it is for, what for, and what must not happen under any circumstances. The clearer the task, the fewer rounds.
Usually yes – but Claude can also be convincingly wrong. Have it show you what it did instead of just believing it.
And whatever can be checked, let Claude check itself: run the tests, redo the arithmetic, name the sources.
Good question – that is exactly where the trap is. Claude reads web pages, e-mails, tickets and other people’s source code. None of that is an instruction from you.
Instructions come from you, from your project rules and from what you allow. Everything else is material to read – which is why the permissions are not there to annoy you.
Glad to hear it.