Claude Code works after installation, and nothing in it needs setting up. So this lesson isn’t fixing something broken – it’s a tour of a place you’ll end up in anyway, so the first visit doesn’t happen in a hurry, in the middle of some problem.
We’ll cover the five settings that matter most. Leave the rest as they are for now – Claude Code’s defaults are considered, and changing something you don’t understand yet gets in the way more often than it helps.
Nothing you do here is irreversible. At the end we’ll change one setting, look at what it did, and put it back – precisely so that going back gets practised too.
Open the settings and look around
Start Claude Code in any folder and type a single slash:
/
A list of commands appears. These are slash commands – instructions for Claude Code itself, as opposed to the ordinary sentences that are the conversation. You don’t have to memorise them or type them out in full: start typing and the list narrows to what matches, the up and down arrows pick the one you want, and Enter runs it. It works like a menu, only summoned from the keyboard.
What we need today is the configuration:
/config
A settings panel opens. You move around it with the up and down arrows and leave with Esc.
One thing to take in straight away, because it saves weeks of guessing: the bottom of the screen always spells out what the keys do right where you are. Usually a value toggles with space or Enter, but not everywhere – effort, for instance, moves with the left and right arrows. Rather than memorising that, glance at the hint line; it always matches whatever you’re standing on.
There are a dozen or so entries, and this is the moment where it’s easy to panic and assume you have to go through all of them. You don’t. Scroll the list once from top to bottom, just to see what’s there, and change nothing yet.
Note the names – the panel is in English, like the rest of the interface. The five we care about today are Auto-compact, Model, Effort, Thinking, and Permission mode.
Check that the conversation summarizes itself
In /config, find Auto-compact and check that it’s on. It is by default, and it should stay that way – we’re looking only so the name isn’t a puzzle later, because sooner or later you’ll see it in a message.
Here’s what it does: Claude Code holds the whole conversation so far in memory, but has limited room for it. When that room starts filling up, instead of simply forgetting the beginning, it summarizes the older part of the conversation and carries on with the summary. A long session doesn’t break off halfway through.
The price is that a summary is a summary – a detail from two hours ago might not survive in it. The practical conclusion for later: anything meant to hold the whole way through is better written into a file than said once in conversation. There’s a lesson of its own about that: your first CLAUDE.md.
Leave Auto-compact on and move along.
See where the model and effort live
Model and effort are the settings you’ll come back to most often – they get changed in the course of ordinary work, not once and for all during first-time setup.
In the panel itself you’ll only see a Model entry. Go into it with space or Enter – both things are inside: you pick the model with Enter, and move effort with the left and right arrows. Effort is set per model, for whichever one is highlighted at that moment – so if you nudge the arrows while sitting on a model other than the one you picked, it’s that other model whose value you just changed.
The model decides how powerful a tool you’re working with. Effort decides how much work the model puts into a single answer before handing it over.
We recommend Sonnet to start with. The default depends on your plan: on the pricier ones it’s Opus – more powerful, but it burns through a great deal more tokens. Excellent for hard problems, but for your first few days it’s a bit like wearing roller skates on a carpet: possible, just not the point. Sonnet is faster, cheaper to run, and more than enough to learn on.
So we’ll change exactly one thing here, and not for good – because the model is the thing you’ll switch more often than anything else in this panel. Which model for what is a separate subject, with a tip of its own.
Since the model changes so often, Claude Code has a shortcut straight to these settings: instead of opening /config and hunting for the Model entry, just type /model. Same list, one command instead of two steps.
And one trap to remember now: a model chosen with /model mid-session saves as the default for the next ones. So switching “just for a moment” isn’t for a moment at all.
Check that thinking is on
The fourth of our five is Thinking, the switch for thinking.
Claude Code can stop before answering and think a task through: work out where to start, what to check, where the catch might be. Thinking decides whether it does that at all.
The easiest way to see it is on yourself. Thinking turned off is the situation where somebody stops you in the street and asks what 8 times 7 is. You’ll probably get it right – but you might also slip, because you’re answering on the spot. Thinking turned on is the same question asked differently: “take your time, and write it out on paper if you need to, before you answer”.
Which is why we recommend leaving Thinking on and forgetting about it. It costs few tokens, and the only thing you really give up is a little time – the answer arrives later, but thought through. You’ll see the difference for yourself at the end of this lesson.
Switch the permission mode and see what Claude Code asks about
The most interesting entry in all of /config is Permission mode, which sets how often Claude Code stops and asks for approval.
The default mode is called Manual and it’s as cautious as it gets: Claude Code reads files without asking, but stops before every change and every command and waits for your confirmation. That’s a good place to start and there’s no hurry to leave it.
Permission mode gets changed often too – you may well switch it several times within one task – so it has a shortcut of its own. Shift+Tab cycles through four modes without opening /config:
- Manual – asks before every edit.
- Accept edits – edits files freely, still asks about commands.
- Plan – researches and proposes a plan, never touching files.
- Auto – Claude Code judges for itself what is safe enough not to ask about.
The current mode shows at the bottom of the screen, so if you overshoot by one, press Shift+Tab again and go round the loop.
Switch to Accept edits and ask Claude Code for something small and reversible:
Create a file called note.txt in this folder and write one sentence into it.
The file appears without a prompt. Now try something else, still in the same mode:
Change the theme to light in my settings file ~/.claude/settings.json.
There Claude Code asks anyway, even though the mode hasn’t changed.
Get the scope of automatic approval exactly right, because it’s easy to overestimate and get more alarmed than the situation deserves: automatic approval covers working files in your folder, not everything. Three places where a mistake hurts most are protected separately, and Claude Code asks about them even in the looser modes. Before you go looking for them: each one is created only when it’s needed, so if you can’t see them in your folder right now, nothing is wrong. The protection applies once they show up.
- .claude – the folder holding Claude Code’s own settings, including everything we were just looking at in
/config. It appears with the first settings saved for a project. - .git – a hidden folder with the project’s change history, created by Git, the program that remembers successive versions of files and lets you go back to them. It appears the first time Git is used in a folder.
- your shell’s startup files – a handful of configuration files your terminal reads every time it opens.
So there’s no accidentally landing in a mode where Claude Code quietly rearranges your system.
After the experiment, go back to Manual if that suits you better. There’s no single right answer here: on work you’re reviewing as you go, Accept edits saves a great deal of clicking, and on something that matters, Manual buys peace of mind. Modes are a big enough subject that we’ll give them a tip of their own – today it’s enough to know you switch them with a shortcut, and that even the loosest of them has limits.
Find the file all of this landed in
The settings from /config don’t live in the cloud or anywhere magical. They’re a plain text file on your disk:
- Windows:
C:\Users\your-name\.claude\settings.json, in your user folder - macOS and Linux:
~/.claude/settings.json
Have a look inside – an ordinary text editor will do. You’ll see a few lines like "effortLevel": "high". Those are exactly the values the panel was showing a moment ago, written out as text.
If the file isn’t there, or holds all of two lines, that isn’t a fault. Only values that have actually been changed get written into it; everything else runs on defaults and doesn’t need spelling out anywhere.
For two reasons. First, that file can be copied to another computer to get the same setup straight away – your login lives separately, so you’ll sign in again on the new machine. Second, when something in your configuration eventually gets tangled, there’s nothing to guess at. Open the file and see what’s actually in it.
There’s a second layer of settings too, belonging to a specific project and saved in a .claude folder inside it. You don’t need it today – it’s enough to know that when a project setting and your own say different things, the one closer to the project wins.
Turn thinking off and see the difference
To finish, an exercise where you change one setting, look at what it did, and put it back. We picked Thinking for it, because it’s the one of today’s five whose effect is visible to the naked eye within a minute.
First, with Thinking on, ask Claude Code something that needs a bit of consideration:
Explain to me why hidden folders exist at all, and why their names start with a dot.
Skim the answer and note how long it was. Then open /config, turn Thinking off, and ask exactly the same question again.
The difference shows immediately: without thinking the answer arrives faster, with thinking it takes longer. What you get for those few seconds is best judged with both answers side by side. That’s knowledge from your own comparison, not from a description in the documentation.
Now turn Thinking back on. The rest of /config stays default, and you’ll come back to it when the way you work changes – usually after a month, when it turns out one permission mode suits you noticeably better than the other. On what’s worth changing there and what to leave alone, there’s a separate tip – no clicking involved, but an explanation of how to recognise that moment.
Tasks
Tick these off as you go. The state is remembered in your browser, so you can come back to this list tomorrow.
-
/configopened and the list scrolled top to bottom - Auto-compact checked and left on
- Model set to Sonnet
- I know that
/modelreaches the same list in one command - I know how Thinking differs from Effort
- Accept edits tried out on a small, reversible change
- Confirmed that Claude Code still asks about a protected file
-
settings.jsonfound and looked at - The same question asked twice, with thinking and without
- Thinking switched back on