If you followed Day 1 you launched Claude with claude --dangerously-skip-permissions. The flag name is alarming on purpose — it’s how flags “you are skipping a safety , are you sure?”. But its actual blast radius is narrower than the name suggests. Here is the precise boundary.
What the flag DOES skip
- Read prompts. Without the flag, Claude asks before reading a file. With the flag, it just reads. (You shared the folder; Claude reading files inside it is the whole point.)
- -command prompts. Without the flag, Claude asks before running
ls,grep,cat, and other read-only . With the flag, it just runs them and you see the output flow past.
What the flag does NOT skip
- Code-change reviews. Every time Claude wants to write or modify a file, it pauses and shows you the diff. You approve each change explicitly. The flag does not change this.
- Destructive-action confirmations.
rm -rf, force-pushes, dropping a table — these still prompt regardless of the flag. - External network calls. Posting to a webhook, calling a paid , opening a browser — still prompts.
In short: the flag mutes the trivial-action dialog, not the meaningful-action dialog.
Why the team uses it as the Day 1 default
Without the flag, every single read and shell command opens a permission prompt. Claude feels glacial. New users bounce in the first hour because the latency feels broken. The flag removes that friction without giving away the reviews that matter.
This isn’t every team’s call. Other teams keep the prompts on and use Shift+Tab to cycle through mid-. Both are reasonable. The NBG AI team’s position is: the noise/signal ratio of the trivial prompts is bad enough that defaulting to the flag is the right tradeoff for the audience this hub serves.
When you might choose differently
- You’re working with a folder you don’t fully trust (e.g. a colleague handed you their ). Drop the flag for that session. Permission prompts give you a chance to see what Claude is about to read.
- You’re auditing Claude’s behaviour for a write-up, a training session, or a security review. Drop the flag — every prompt becomes a visible decision.
- You’re paranoid for any reason at all. Drop the flag. Latency-pain is real but knowable; not knowing what your tool is doing is worse.
To launch without the flag: just claude (no flag). You can also press Shift+Tab within a session to change permission mode — see the permission-modes tip.
There’s a middle setting most people skip past
The flag is a blunt instrument, and reaching for it is usually a reaction to prompt fatigue rather than a considered choice. Two better tools exist:
/fewer-permission-prompts— pre-approves the safe, read-only commands you already keep approving, based on your own usage. Removes most of the noise without removing the reviews./permissions— lets you write deny rules, which hold no matter which mode you’re in. A deny rule on.envreads orgit pushstays in force even on the days you’re moving fast.
That second one is the important half for bank work: a deny rule is a control you can point at and describe. “I’m careful” isn’t. See the permission-rules-not-blanket-trust tip.
Also worth knowing: the default permission mode is now auto, where Claude vets each action for risk and prompt injection before running it and only asks about what warrants asking. Much of the prompt fatigue this flag was invented to solve is less severe than it was.
The reconciliation you might be looking for: the “always review the diff” advice you see elsewhere on this hub refers to code-change reviews — which this flag does not touch. The flag and the always-review-the-diff habit live at different layers of the permission system. They coexist by design.