Configuring Claude

Once Claude Code is installed, a handful of configurations make it productive for daily work. The most important is auto mode; everything else is nice-to-have.

Enable auto mode

Out of the box, Claude Code asks you to approve every potentially-sensitive action: file writes, bash commands, network calls. Safe, but unworkable in practice. A real session has dozens of those per minute; every prompt is a full stop, and the model’s momentum dies between actions.

Three permission modes exist:

• Default (approve-per-action): what you get on install. Interrupt-heavy. Fine for your first ten minutes; not a working mode.
• Auto mode: Claude runs in a constrained sandbox. Pre-approved tool categories skip the prompt; genuinely dangerous actions still get stopped. This is where you should be most of the time.
• Dangerously skip permissions: no prompts, ever. Fast. Reserved for sandboxed environments (containers, VMs, scratch repos) where Claude can’t break anything you care about.

The dance every user has to do is turning auto mode on. Three ways, in order of durability:

For a single session

claude --permission-mode auto

For every claude invocation in your shell

Add an alias to your shell config. For zsh:

# ~/.zshrc
alias claude='claude --permission-mode auto'

For bash, the same line in ~/.bashrc. Reload (source ~/.zshrc) and claude runs in auto mode from then on.

As Claude Code’s own default

Better: set it in the Claude Code config so it applies regardless of how you launch it.

Edit ~/.claude/settings.json:

{
  "permissions": {
    "defaultMode": "auto"
  }
}

You can use "bypassPermissions" for full yolo, though I don’t recommend it as a default.

See Modes and abilities for the full permission model and when each level is appropriate. This chapter is the minimum to be productive; that chapter is the full context.

The basic tour

A handful of commands worth knowing on day one.

• Slash commands: type / to see the built-in commands. Tab-completes. Examples:
  • /usage: where you stand against your plan’s limits
  • /model: which model you’re on, and switch between them
• Invoke by intent: instead of typing a slash command, describe what you want and Claude fuzzy-matches to a relevant Skill or command. Explicit /commands are faster once you know them; intent works fine when you don’t.
• Models and reasoning effort: via /model you can switch between Opus, Sonnet, and Haiku variants. Newer versions (Claude Fable 5, Opus 4.8, Sonnet 5) generally mean better outputs. You can also set a reasoning effort: low, medium, or high. Higher effort spends more reasoning tokens reflecting before responding: slower and more expensive, but outputs are more considered.
• Exiting: press Ctrl-C twice to interrupt and exit. Type exit. Or just close the terminal tab. Claude Code isn’t a long-running process that needs a clean shutdown.

A note on model choice and reasoning effort

The gap between successive generations, or between low and high reasoning effort, is real, but smaller than you’d expect. The outcome difference is dominated by context.

The models are already more capable than most users know how to drive. Feed the model a clean, congruent set of files and a well-framed prompt and even a cheaper model produces excellent work. Feed a frontier model a chaotic mix of outdated notes, half-finished plans, and conflicting instructions, and it still struggles, because you’ve asked it to average the chaos.

Most of this course is about getting context right. The model picker and reasoning-effort knobs are a distant second concern.

Other configuration worth knowing early

• The Claude Chrome extension: gives Claude a real browser via MCP. Install from the Chrome Web Store if browser control is part of what you’ll do (see Browser control).
• Other Claude Code surfaces: Claude Code also ships as a desktop app, a web app (claude.ai/code), and IDE extensions, if the terminal isn’t your preferred surface. The separate Claude.ai app is for chat and voice.
• VS Code: a full code editor if you want to review or edit code directly. Has a built-in terminal (Ctrl-` ) so you can run claude right inside it, and can open a folder to see Claude’s file changes as they happen. Use it for review, not for typing.
• Voice dictation: Wispr Flow streams your voice into Claude Code with excellent transcription. Install it, then double-tap the globe (🌐) key on your Mac to start dictation; double-tap again to stop, and the transcribed text drops into your cursor. There’s a free trial and a monthly plan; check the current pricing on their site. Claude has built-in voice, but Wispr is still better for hands-on-keyboard work. See Design as artifact for how voice changes the workflow.
• Claude Code IDE plug-ins: official integrations for VS Code and JetBrains that keep your editor synced with what Claude is doing.

None of these are required. Add them as the course calls for them.

Resources

• Anthropic: Claude Code settings
• Anthropic: Claude Code reference