Claude Code’s settings panel opens with /config and looks exactly the way settings panels look: a column of names, a column of values, and not one sentence about what happens when you flip any of them. Some entries you can decode from the name alone. Others read like the name of a feature you’re not sure you even use.
So we’re going through the ones that matter – not in the order they sit on screen, but in four groups, by what each one actually affects. The ones you’ll notice on your token usage come first, because that’s where a wrong guess costs the most.
One thing about the panel itself before we start. Along the bottom of the screen, Claude Code prints what the keys do in the option you have selected, and that line changes from option to option – a plain toggle gets different keys than an option with a list of values to pick from. When you can’t work out how to change what you have highlighted, the answer is at the bottom of the screen.
We’ve written a separate lesson on the settings that matter most – this post goes deeper. And one caveat before the details: Claude Code moves fast. An entry that’s called one thing today may be called something else next month, and new ones turn up beside it.
What drives your token usage
Five entries where the setting makes a difference to how fast you hit your plan’s limit. The last two can change usage by an order of magnitude, so read those even if you change nothing.
Auto-compact
Claude Code holds your whole conversation in memory, but the room for it is finite. With auto-compact on, the older part of the conversation gets summarised at the right moment and work carries on from the summary; with it off, the conversation reaches the ceiling and stops. Leave it on – that’s the default. Just keep in mind what follows from it: a detail you typed two hours ago may not survive the summary, so anything meant to hold at all times belongs in a file of instructions.
Thinking mode
With thinking mode on, Claude spends part of its work on reasoning that never appears in the answer itself. It’s on by default and should stay that way – we’ve covered it at length in the tip about models, alongside the effort level it often gets confused with.
Model
You can pick the model in the settings panel, or jump straight to its own screen with /model; the default value matches the model to your plan. We won’t give you a one-line recommendation here, because the right choice depends on the kind of work – we’ve covered it in separate posts: the tip on which model for what and the lesson on first-run settings.
Dynamic workflows and Dynamic workflow size
With Dynamic workflows on, Claude Code can respond to a harder task by writing itself a script that splits the work between many helper agents running in parallel, then collects their results – on a large search, or a review across many files at once, that makes a real difference to how long it takes. The parallelism has a price, and Claude Code says so outright: many agents working at once can burn a great deal of tokens in a short time, and all of it counts against your plan’s limit. We’d turn it on, but knowingly – and start by checking where this entry stands on your machine, because the default depends on your plan. A running workflow can be stopped at any point with /workflows.
The entry below it, Dynamic workflow size, sets the upper number of agents in one such workflow: small means fewer than five, medium fewer than fifteen, while large and unrestricted go considerably higher. Medium is the default – leave it there, or drop to small if you’d rather a workflow got a handful of agents instead of a dozen and a half.
Ultracode keyword trigger
Ultracode is a mode where Claude Code works at the highest effort level and treats orchestrating many agents as its default approach rather than a last resort; it needs Dynamic workflows on and a model that supports the highest effort. The entry in the panel, though, switches something narrower than the name suggests: it decides whether the word “ultracode” appearing in your message should turn that mode on automatically for that one answer, and by default it does. We’d turn it off – not because the mode is bad, since it’s excellent for a big one-off review, but because triggering it with a stray word in a sentence is the most expensive way imaginable to start it. With the trigger off the mode is still there: just ask for it directly.
How much control stays with you
The entries in this group decide what Claude Code asks about before it acts, where the results of its work end up, and where the version running on your machine comes from.
Default permission mode
The permission mode decides what Claude Code asks about before doing anything. Reading files needs no approval in any mode – the differences start with changes and commands.
- Manual – the default. Every file change and every command waits for your confirmation.
- Accept edits – file changes inside the folder you’re working in, plus a handful of basic file operations, go through without asking.
- Auto – a mode that assesses each action and decides for itself whether it needs to ask. In practice Claude Code edits files, runs scripts and makes small changes to your system on its own, and stops only at more serious interventions.
- Don’t ask – Claude Code asks nothing at all, and that doesn’t mean it’s allowed more. The opposite: anything it would have to ask about is automatically refused.
- Plan – a mode where Claude Code executes nothing. It reads, analyses and lays out a plan of action, and you decide whether to approve it. Once you approve a plan, the mode switches automatically to Accept edits.
Don’t ask is worth a moment longer, because the name says something different from the behaviour. The only things that get through are those already covered by an approval written into your permission rules, plus reading files. Everything else Claude Code refuses on its own, without asking and without waiting. The approval has to exist before it even tries. Approvals you click during work count only where you picked the “don’t ask again” option – those go into the rules and apply here too. A plain “yes” in Accept edits covers one situation and disappears with it. To see what’s already in your rules, run /permissions. The mode was built for runs where nobody is sitting at the keyboard, and there’s no reason to switch it on for ordinary work.
For a start we’d stay on Manual. When a few weeks in it turns out that Claude Code does roughly what you ask of it, and that you’re approving its changes on reflex, moving to Auto or Accept edits gives it more independence without handing over control blind. Plan works differently from the others, so treat it separately: you don’t set it once and forget it, you switch it on when a task is bigger and you want to see what Claude Code intends to do first.
One thing here holds regardless of mode – automatic approval stops at working files, because the directory with your version history, the directory with Claude Code’s own settings, your shell’s startup files and a few other places are protected separately, and writing to them needs approval even in the looser modes.
Question auto-continue timeout
When Claude Code asks you a question with options to choose from, it stops until you answer. Question auto-continue timeout lets you set how long it waits before it stops waiting and picks the first, recommended option itself. The choices are a minute, five minutes, ten minutes or “never”, and the last one is the default. Leave it on “never” at first, because until you know what Claude Code tends to ask about, you don’t want it answering on your behalf. Later it becomes useful in a fairly specific way: if you leave Claude Code with a longer task and walk away from the computer, five or ten minutes means you won’t come back to find it parked on one question for an hour.
Artifacts
An artifact is a finished thing – a page, a small interface, a visualisation – that Claude publishes at its own address so you can send someone a link to it. We’d turn it off, and for a specific reason: one of the better things about working in Claude Code is that everything you make is made on your own disk and stays there even after you stop using the tool, and an artifact changes that ownership, because the file ends up on Claude’s side. If you happen to be making something to show the outside world, turn it on for that one occasion.
Auto-update channel
Claude Code updates itself, at launch, and this entry only decides where it takes versions from: latest means every new release straight away, stable means versions that have been out for a while. Pick stable – new releases come often and sometimes something in them doesn’t work, and the stable channel means such a version has time to get fixed before it reaches you. Switching itself brings up a question, because the stable channel is sometimes older than the version you’re on: Claude Code will ask whether to allow rolling back to the stable channel’s version, or to stay on the current one until the channel catches up. On that one question, pick the second answer – you keep what already works, and it won’t come up again.
Enable Remote Control for all sessions
Remote control lets you look in on a Claude Code session and steer it from outside the terminal it started in, and this entry decides one thing: whether that connection should establish itself at every launch. True means yes, every session starts with a connection ready; false means none do; default just follows whatever is set on your account. We’re not pushing a choice here – it depends on something invisible from the outside, namely whether you use remote control at all. Default and false are two different values, and that’s easy to miss.
How Claude Code communicates
Four entries that don’t change what Claude Code does – only how much it says about it, and how.
Output style
The manner Claude works in. Five to choose from:
- Default – ordinary work, no extra roles.
- Explanatory – Claude explains its choices as it goes: why it did something a particular way, how the thing it just used works.
- Learning – study mode. Claude stops and asks you to write small pieces yourself, to give you practice instead of a finished result.
- Proactive – Claude acts immediately, interrupts with questions less, and would rather do than plan.
- Concise – Claude leads with the result, skips the preamble and the account of what it just did, and keeps answers short. The work itself gets done as thoroughly as it does in Default. Ask for an explanation and you get it in full; error reports, security warnings and questions about irreversible actions stay complete either way. This one arrived in August 2026, so update Claude Code if you don’t see it on the list.
We’d go with Default or Explanatory – the latter when you want to understand what’s happening along the way and don’t mind a little more text. Learning only makes sense if you’re using Claude to learn: in that mode Claude takes on a teacher’s role and sends part of the work back to you, which isn’t much use when what you want is tasks automated and Claude as autonomous as possible. Concise is for later, once you have worked with Claude Code long enough that its running commentary tells you nothing new and you just want the result. Early on the commentary is worth keeping: it is how you learn what Claude Code is actually doing to your files. Proactive can be very comfortable, but only once the working relationship is established and you know what to expect from Claude. Before that, the style works against you: since Claude interrupts with questions less often, you get fewer moments to turn it around before it does something wrong.
Verbose output
Verbose mode on shows the full output of the tools Claude Code uses – the entire contents of a file it read, the entire response from a command – while with it off you see shortened summaries. Better left off, because full output turns the terminal into a wall of text you have to scroll through to find a sentence from Claude. This setting only changes how much the tools Claude uses give back – how talkative the assistant itself is lives in Output style above, or in your file of instructions. We’d turn verbose on only for debugging, when something isn’t working and you want to see exactly what a command returned.
Prompt suggestions
In the field where you type, a ready-made suggestion appears of what you might write – picked to fit what just happened in the conversation. Claude Code asks whether it should do something, and “yes, go ahead” is already waiting in the field. You accept a suggestion with Tab.
If you don’t want the suggested message, you simply start writing your own in the reply box. We’d leave this on: early on it’s one of the few places where you can see how requests to Claude Code are phrased at all.
Language
The only entry in the panel where you type the value in by hand – you give the language name in English or as a two-letter code, for example polish or pl, and typing default returns to English. It doesn’t change the interface: Claude Code’s labels, commands and messages stay in English and there’s no setting that changes that. When you’re typing, this entry changes little, because the conversation follows whatever language you ask your questions in – it makes a real difference when you dictate by voice, where there’s no text to recognise a language from.
Looks and comfort
Five entries that change nothing beyond what you see on screen. None of them can break anything except your own comfort.
Theme
The interface’s colour theme. We’re only noting that this is where you change it.
Show tips
Hints shown next to the working indicator while Claude Code is busy – short lines along the lines of “you can drag a file into the terminal window”. We’ve no recommendation here, since it’s a matter of taste: some people enjoy them, others find them distracting. We do have one correction, because it comes up in training again and again – a tip means “this feature exists”, not “something is wrong and you need to react”. A hint appears at a random moment and bears no relation to what you’re doing, so nobody needs to click or check anything because of it. That distinction is worth keeping in mind especially in your first week, when every new message in the terminal looks like something that wants attention.
Reduce motion
Cuts down interface animation – the pulsing of the working indicator, effects when the screen changes, that sort of thing. Off by default; turn it on if animation bothers you.
Terminal progress bar
During longer operations Claude Code sends progress information to the terminal, and how that gets displayed is up to the terminal: one will draw a bar in the window title, another will light up the icon in the taskbar, a third will ignore it completely. Turn it on – when it works, you get a signal that something is happening without looking at the terminal window, and when your terminal doesn’t support it, nothing happens at all.
Show turn duration
After each answer Claude Code notes how long it took to prepare. We’d turn this on too: it’s the cheapest way to develop a feel for what’s fast in this tool and what isn’t – after a few days you start reading things out of those numbers that aren’t visible otherwise, such as that a question about one file comes back at once while searching a whole project takes forever.
In short
There are only a few changes genuinely worth making the first time you open the panel, and they all fit inside a minute. Turn off the “ultracode” keyword trigger and Artifacts, check that Dynamic workflow size is on small or medium, switch the update channel to stable, and on the version question, stay on the current one. If you like feedback from your terminal, turn on the progress bar and the turn duration as well. The rest can stay as it is.
You’ll come back to two entries later – that’s normal: the permission mode will stop fitting on the day you start leaving Claude with a task instead of leading it by the hand, and Question auto-continue timeout will earn its place the first day you walk away from the computer while it works. The rest of the panel has sensible defaults – most likely you won’t need to change them.