Claude Code Skip Permission Confirmation

Originally written as internal engineering notes while building the thing it describes. Republished here with light editing — the timeframe is noted where it matters.

Problem#

During Claude Code usage, frequent permission confirmation prompts (e.g., Bash command execution, file editing) require manually selecting "Yes", which interrupts the development workflow.

Solution#

Temporary (Current Session Only)#

# Option 1: Dedicated flag, same effect as Option 2
claude --dangerously-skip-permissions
 
# Option 2: General parameter, bypassPermissions is one of the available values
claude --permission-mode bypassPermissions

Both options have the same effect — they skip all permission prompts. The only difference is the entry point: one is a shortcut flag, the other is a specific value of a general parameter.

Persistent, Fine-Grained (Recommended)#

Rather than switching the checks off, pre-authorize the tools you actually get asked about, in ~/.claude/settings.json:

{
  "permissions": {
    "defaultMode": "acceptEdits",
    "allow": [
      "Bash(*)",
      "Read",
      "Write",
      "Edit"
    ]
  }
}

acceptEdits automatically accepts file edits, and Bash(*) covers Bash command confirmations. Everything not on the allow list still prompts — which is the point: the prompts you kept are the ones for things you didn't anticipate.

Narrow the entries further as you learn your own patterns. Bash(git *) or Bash(npm run *) removes most of the noise while leaving an arbitrary shell command as something you still get to see before it runs.

Persistent, Global Bypass (If You Accept the Risk)#

The blunt version — applies to every project on the machine:

{
  "permissions": {
    "defaultMode": "bypassPermissions"
  }
}

Worth being clear about what this turns off. The permission prompt is the only place where a command is shown to you before it executes — so bypassing it means file writes outside the project, rm, curl to an arbitrary host, git push, and anything an agent picked up from a file it read all run without you seeing them first. That last case is the one to weigh: an agent acting on instructions embedded in a repository, a web page, or a dependency's README is exactly the scenario the prompt exists to catch.

I use it only in throwaway containers and sandboxes. On a machine holding real credentials or a real repo, the fine-grained setup above is the better trade.

permission-mode Available Values#

Summary#

Start with the allow list. It removes the prompts that are pure friction — the ones you would have approved without reading — and keeps the ones that are actually doing a job. bypassPermissions skips all of the checks, and is only a sensible default in an environment you would be willing to throw away.