Master
AI-Powered
Coding Tools

Essential Commands

Permission Modes

Claude Code asks permission before running commands or editing files. Three built-in modes:

Start with Ask mode. Graduate to auto-edit when you trust the workflow.

What Claude Code Does

• Reads and understands your entire codebase
• Edits multiple files in coordinated changes
• Runs shell commands, tests, and builds
• Creates commits and pull requests
• Connects external tools via MCP servers
• Runs custom agent teams (subagents)
• Works in Terminal, VS Code, JetBrains, Desktop, Web

Best Practices

• Be specific in your requests — "fix the login bug in auth.ts" beats "fix bugs"
• Let Claude explore first — "what does this codebase do?" before asking for changes
• Use /compact in long sessions to free context space
• Write a CLAUDE.md so Claude knows project conventions from the start
• Review changes before committing — Claude is good, but verify

Where to Define Hooks

How Hooks Respond

Or return JSON on exit 0 for richer control:

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "deny",
    "permissionDecisionReason": "Destructive command blocked"
  }
}

Block Destructive Commands

A PreToolUse hook that prevents rm -rf:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{
        "type": "command",
        "if": "Bash(rm *)",
        "command": ".claude/hooks/block-rm.sh"
      }]
    }]
  }
}

The script reads JSON stdin, checks the command, and returns permissionDecision: "deny" if it contains rm -rf.

Background & Prompt Hooks

• Async hooks — Add "async": true to run in background while Claude continues
• Prompt hooks — LLM evaluates conditions with yes/no JSON response
• Agent hooks — Subagent with tool access verifies before deciding
• HTTP hooks — POST to a URL, response body for decisions
• PowerShell — Set "shell": "powershell" for Windows hooks
• MCP matching — Use mcp__server__tool patterns in matchers
• disableAllHooks — Temporarily turn off all hooks in settings
• Run claude --debug to see hook execution details

Gemma 4 Model Sizes

E4B is the sweet spot for most developers. 26B uses Mixture-of-Experts with only 3.8B active params.

What Makes Gemma 4 Special

• Multimodal — text, image, and audio input (E2B/E4B)
• Thinking mode — configurable reasoning with <|think|>
• Native tool calling — function calling for agentic workflows
• System prompts — native system role support
• Variable resolution — token budgets for image detail control
• 128K–256K context — process long documents locally

Recommended Settings

For best results with Gemma 4:

For thinking mode, include <|think|> at the start of the system prompt. In multi-turn conversations, strip thoughts from history — only keep final responses.