Master Claude Code CLI for Max Productivity

A deep specialist in Claude Code, Anthropic's CLI, covering shortcuts, hooks, MCP servers, advanced configuration, workflows, CLAUDE.md, memory, sub-agents, permissions, and ecosystem integration. Unlike the Claude.ai web interface, Claude Code accesses the filesystem directly, runs bash/git/npm commands, persists context via CLAUDE.md and memory files, supports MCP server tool extensions, supports pre- and post-action hooks, and can spawn and orchestrate sub-agents via the Task tool.

Use this when configuring Claude Code, creating or debugging hooks, adding or configuring MCP servers, creating sub-agents and parallel orchestration, optimizing CLAUDE.md and memory files, configuring permissions and security, or resolving unexpected CLI errors or behavior.

Not needed for tasks unrelated to Claude Code itself, or when a simpler, more specific tool already handles the request.

Covers installation and core flags: npm install -g @anthropic-ai/claude-code, then claude for an interactive session or claude "task" for non-interactive mode. Key flags include -p for print/script mode, --model, --max-tokens, --output-format json, --allowed-tools, --max-turns, and --dangerously-skip-permissions - reserved for controlled CI/CD, automated scripts, or isolated sandboxes, and never used in production or shared environments holding secrets.

CLAUDE.md is loaded automatically every session at three levels - a global ~/.claude/CLAUDE.md, project root, and subfolder - and its recommended structure covers project context, essential commands, code conventions, architecture, critical business rules, available agents and skills, and a pre-task protocol. Memory files live under ~/.claude/projects/<path-hash>/memory/, with MEMORY.md as a roughly 200-line index plus per-topic files for personas, project context, and decisions; a context-guardian pattern monitors auto-compaction and saves snapshots for long or critical sessions.

Hooks, configured in ~/.claude/settings.json globally or .claude/settings.json per project, fire on PreToolUse, PostToolUse, Notification, Stop, and SubagentStop events - covering use cases like a completion sound, logging every bash action, or a pre-commit secret scanner, verified afterward by inspecting and JSON-validating the settings file. MCP servers extend Claude Code with external tools via claude mcp add/list/get/remove; commonly used ones include filesystem, github, postgres/sqlite, puppeteer/playwright, slack, and fetch, and a custom MCP server can be built in Node.js with the @modelcontextprotocol/sdk and registered the same way. Sub-agents spawn via the Task tool for parallel work - general-purpose, Bash, Explore, and Plan types - ideally run in the background for long tasks, given full context explicitly since sub-agents don't see conversation history, and optionally isolated on their own git worktree for risky experiments. A small set of environment variables control the CLI directly: ANTHROPIC_API_KEY for authentication, CLAUDE_CODE_MAX_OUTPUT_TOKENS to raise the output token ceiling, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC for a more private mode. Permissions are configured per project with allow/deny glob patterns, such as allowing Bash(git *) while denying Bash(rm -rf *), or via --allowed-tools/--disallowed-tools flags. A complete feature workflow runs four phases - planning, implementation, testing, and code review - each as a separate invocation, and slash commands (/status, /clear, /compact, /memory, /hooks, /mcp, /cost, /model, /help) manage the session interactively.

Developers configuring or troubleshooting Claude Code - setting up hooks and MCP servers, writing an effective CLAUDE.md, orchestrating sub-agents, or tuning permissions and security - who want concrete commands and configuration guidance rather than general CLI documentation.