AI CODING TOOLS / LESSON 05

Permission Modes.

Controlling what Claude Code can do

Written by Andrew Lee JenkinsPublished Oct 5, 202610 min readFree lesson

Claude Code can create files, edit code, run commands and make REAL changes to your computer. That's awesome... and it's also exactly why you need to stay in the driver's seat. Permission modes are how you decide how much freedom Claude gets.

Why Permissions Matter

Say you hire a contractor to work on your house. You've got options.

Stand there and approve every single nail they hammer (super safe, super slow)
Let them work while you check in now and then (faster, a little riskier)
Hand them a detailed plan and trust them to run with it (efficient, if you trust them)

Permission modes work the same way. You pick how much oversight you want.

The Three Permission Modes

Claude Code has three main modes for handling permissions. (There are a few more advanced ones too, but these are the three you'll touch first.)

1. Ask Mode (Manual)

In ask mode, Claude stops and asks before doing anything significant. It's the safest option. Claude Code calls it Manual mode these days, and newer versions (v2.1.283 and up) actually start sessions in auto mode, where a separate safety checker reviews actions for you. Press Shift+Tab during a session to cycle modes and land on manual.

How it works.

Claude explains what it wants to do
You see the exact file or command
You approve or reject each action

Example.

Example

Create file
index.html
[shows the code]

Do you want to proceed?
❯ 1. Yes
  2. Yes, and don't ask again (exact wording varies)
  3. No

Pick an option with the arrow keys and Enter (or just press the number). Pressing Esc is the same as saying No.

2. Auto-Accept Mode

In auto-accept mode, Claude makes changes without checking in every time. It just does the work and tells you what it did.

How it works.

Claude makes changes immediately
You see a summary of what changed
You can still stop it or undo changes

When to use it.

When you trust Claude with a specific task
When you want to move faster
When you're doing repetitive work

To enable it. Press Shift+Tab during a session until the status bar says "accept edits on". That lets file edits through without asking. The --dangerously-skip-permissions flag goes a LOT further. It's bypass permissions mode, which skips every permission prompt for everything, commands included.

Example

claude --dangerously-skip-permissions

3. Plan Mode

Plan mode sits in the middle. Claude builds a complete plan, you review and approve the whole thing, and then it gets to work.

How it works.

Claude analyzes your request
It creates a step-by-step plan
You review the entire plan at once
Once approved, Claude executes all steps

When to use it.

For complex tasks with lots of steps
When you want to review the overall approach
When you trust the plan but not every lil micro-decision

Choosing the Right Mode

Here's a quick cheat sheet.

Situation	Recommended Mode
Learning Claude Code	Ask Mode
Important project files	Ask Mode
Quick prototyping	Auto-Accept Mode
Large refactoring task	Plan Mode
Running unfamiliar commands	Ask Mode
Repetitive file edits	Auto-Accept Mode

What Actions Need Permission?

Plenty of stuff doesn't need a thumbs up. Here's what Claude actually asks about.

Always asks in ask mode.

Creating new files
Editing existing files
Deleting files
Running terminal commands
Installing packages

Doesn't need permission.

Reading files (just looking, not changing)
Explaining code
Answering questions
Showing you information

Saying No

You can ALWAYS say no. If Claude proposes something that looks wrong, here's the drill.

Pick No (or press Esc) to reject the action
Explain what you'd rather it do
Claude adjusts and tries a different approach

Example.

Example

Claude wants to run: rm config.js
Do you want to proceed?  → No

You: Actually, let's keep that file but rename it to config.backup.js

Claude: Good idea! I'll rename config.js to config.backup.js instead.

Permission Prompts Explained

When Claude asks permission, the prompt usually shows you three things, plus your answer options.

What it wants to do - "Create file" or "Run command"
Where/what - The filename or command
The content - What it will write or the full command

Something like this.

Example

Create file: /projects/mysite/index.html

<!DOCTYPE html>
<html>
  <head>
    <title>My Site</title>
  </head>
  <body>
    <h1>Hello World</h1>
  </body>
</html>

Do you want to proceed?
❯ 1. Yes
  2. Yes, and don't ask again (exact wording varies)
  3. No

The middle option is handy. It says yes AND remembers your answer. For file edits that lasts until the session ends, and for a terminal command it's saved for that project, so Claude stops asking about that same command.

You know exactly what's happening before it happens.

A Note on Safety

Claude Code is built to be safe, and even in auto-accept mode it tries to steer clear of dangerous actions. But no system is perfect (anybody who tells you theirs is, I'd keep a hand on my wallet). A few safety tips.

TL;DR

Permission modes control how much freedom Claude has
Ask mode (Manual) is the safest, since you approve each action
Auto-accept mode is faster but riskier
Plan mode lets you review a complete plan before anything runs
You can always say no and suggest a different approach

What's Next?

Now that you know how to keep Claude in check, let's learn how to talk to it so you get great results. Next lesson is all about writing prompts that actually land.

Loading the lesson…

Loading activities…