A permission mode sets which Claude Code actions need approval. Choose one based on the work and the risk you are willing to accept. section 5.6
The available modes trade approval checks for autonomy:
- manual pre-approves file reads and read-only commands; other actions require approval. section 5.6
- acceptEdits pre-approves reads, edits, and commands judged safe within the working directory. section 5.6
- plan allows reads and navigation, but not project edits; Claude writes a plan for you to approve. section 5.6
- auto replaces the human approver with an AI classifier that judges permissions. section 5.6
- bypassPermissions skips permission checks except for commands that would wipe the hard drive. section 5.6
Treat bypassPermissions as a risk decision, not a convenience default: malicious instructions in a cloned repository could prompt dangerous commands. It may be acceptable for a prototype using your own code in a dedicated environment that cannot affect other systems. section 5.6
Plan before building
For a messaging feature, plan mode can help shape a strategy before implementation. It permits read-only work and navigation, and can write the plan as a Markdown file outside the project repository. Review and approve the plan before Claude starts work. section 6.3
This command requires approval Do you want to proceed? > 1. Yes 2. Yes, and allow access to caw-project/ and git ls-tree commands 3. No
An approval prompt shows the choice between approving a single action, allowing a broader set of actions, or declining. section 5.6
claude --permission-mode auto /plan
The listing shows a command-line mode flag and the plan command; plan, auto, and manual modes can also be cycled with Shift-Tab. section 6.3
