What you will learn The four autonomy levels you choose when handing over a task, and which situation each suits.
When you direct a task you also choose the width of its autonomy.
| Choice | How it behaves |
|---|---|
| Fully automatic (default) | runs to the end without stopping to ask |
| Auto-approve file edits | file changes proceed unasked; everything else confirms as needed |
| Plan only | shows you a plan and does not execute |
| Standard | Claude Code's own default behaviour |
It is safe because nothing executes. It also produces nothing.
File changes happen on their own; anything else (running commands and so on) asks for confirmation. It draws the line between what can be undone and what cannot.
If you use git, file changes are generally reversible. That is what makes this level the balance point between safety and convenience.
It is not the default, so you have to select it. See the box below.
If you choose nothing when creating a task, you get this level. It runs to the end without stopping to ask, so for a folder you are new to, or work that is hard to undo, always change it to something else.
Exactly how you have always used Claude Code. For people who are already comfortable with it.
HyperTeamsgraph TD
A["Plan only<br/>run it a few times"] --> B["Do the results match expectations?"]
B -->|"yes"| C["Auto-approve file edits"]
B -->|"no"| D["Fix the instruction"]
D --> A
C --> E["Have you done this type several times?"]
E -->|"yes"| F["Fully automatic"]Do not go straight to fully automatic. Checking the results at each step and climbing is faster — because there is no incident to clean up.
This is why splitting working directories by risk was suggested.
| Folder | Default autonomy |
|---|---|
| Experimental / personal projects | fully automatic |
| Work projects | auto-approve file edits |
| Anywhere with production deployment scripts | plan only |
Set a different default per folder and the right level applies without you choosing every time.
Raising autonomy and having safety nets are two different things.
Safety nets look like this:
With safety nets you can raise autonomy, because you can undo.
Fast and good are different. Set fully automatic on a type of work you are doing for the first time and the agent goes just as hard in the wrong direction. With no checkpoint, you find out later.
The time spent reading the plan and fixing the instruction is shorter than the time spent undoing something executed wrongly. Especially for the first few runs.
1. Which level do you get if you choose nothing?
Fully automatic. It runs to the end without stopping to ask, so for a folder you are new to or work that is hard to undo you must change it. The balance point between safety and convenience is auto-approve file edits, but you have to select that yourself.
2. How should you climb through autonomy levels?
Plan only → auto-approve file edits → fully automatic, checking at each step that the results match expectations. Going straight to fully automatic costs more in incident cleanup.
3. Under what condition is it fine to raise autonomy?
When safety nets exist — commit before starting, a narrow working directory, production data excluded, the habit of checking results. If you can undo it, you can raise autonomy.
Next, the always-open command window → The terminal
Use when: □ it's your first time working in this folder □ you have no feel for what it will touch □ the work is hard to undo (migrations, bulk edits) □ you have to show someone else the planUse when: □ you've done this type of work several times with stable results □ nobody is watching, as with scheduled runs □ failure has small consequencesDo not use when: □ it's a type of work you're doing for the first time □ it affects production □ it's hard to undoHigh autonomy + safety nets → fast and safeHigh autonomy + no safety nets → dangerousLow autonomy + no safety nets → slow and anxious□ commit before starting□ keep the working directory narrow□ never register a folder holding production data□ the habit of checking results