Master Claude Code: Full Guide to AI-Powered Programming, Permissions & Agent Teams

Claude Code Overview

Claude Code is an AI programming agent from Anthropic, built on the Claude large model to strengthen ecosystem competitiveness. It can write, edit, and debug code, translate and polish text, and perform data analysis, Excel handling, PDF processing, and information summarization.

Project Initialization

A Project represents a folder or repository. To let Claude Code manage it, run the /init command, which creates a CLAUDE.md file containing the project’s specifications and requirements, providing persistent context for future sessions.

The CLAUDE.md file can be edited manually or via the /memory command. It has two scopes: User Scope (under the home directory) and Project Scope (inside the project folder), with Project Scope taking precedence.

Permission Modes

Claude Code supports three file‑editing permission modes, switched with Shift+Tab :

Default mode : Claude asks before editing. Options are Yes (allow once), No (deny once), or Yes + allow all edits for the session (Shift+Tab).

Edit mode : Higher‑privilege automatic editing, suited for coding tasks.

Plan mode : Claude analyses code without modifying files, discusses a plan, then offers four choices: auto‑accept with context clear, auto‑accept without clearing, manual approval each edit, or continue dialogue to refine the plan.

For full automation, start Claude Code with --dangerously-skip-permissions, but this is risky and should only be used in sandbox environments.

Session and Context Management

Each Project opens a Claude Code session that stores conversation history. New sessions start without previous history. To resume a prior session, use claude -c or the /resume command. Sessions can be renamed with /rename and cleared with /clear.

When the context window approaches 95% capacity, Claude automatically compresses it. You can manually trigger compression with /compact. If the LLM exceeds its context limit, manual compression may be required.

Code Version Control

After issuing a development request, Claude edits files and negotiates permissions. If the changes are unsatisfactory, use /rewind to roll back, choosing among restoring code and conversation, restoring conversation only, restoring code only, or cancelling the rollback. Note that /rewind only affects code generated by Claude, not user‑written code.

Shell Mode

Commands prefixed with ! are executed as shell commands, preserving the familiar terminal experience.

Hooks

Hooks are JSON‑defined triggers that run automatically on specific events such as SessionStart , SessionEnd , or PreToolUse . Use /hooks to list available hooks. Example: a Notification hook that runs

osascript -e 'display notification "Claude Code needs your attention" with title "Claude Code"'

on macOS.

Hooks can be stored in Project Scope (local, not git‑tracked), Project Scope (git‑tracked), or User Scope.

MCP (Multi‑Channel Protocol)

Claude Code acts as an MCP host, allowing installation of MCP servers and clients, such as the Figma MCP:

claude mcp add --transport http figma https://mcp.figma.com/mcp

After installation, use /mcp to view status and configure API keys.

AgentSkills

Installed Skills can be listed with /skills. They extend Claude Code’s capabilities.

Subagents

Subagents are specialized agents for tasks like code quality improvement. Create a subagent by adding a markdown file under ~/.claude/agents:

---
name: code-quality-improver
description: Reviews code for quality and best practices
tools: Read, Glob, Grep
model: sonnet
---
You are a code quality improver. When invoked, analyze the code and provide specific, actionable feedback on quality, security, and best practices.

List subagents with /agents and invoke them via natural language.

Plugins

Plugins package Skills, Subagents, MCP, and Hooks for distribution. The official Plugin marketplace provides additional functionality.

AgentTeams

AgentTeams enable multi‑agent orchestration. A Team Leader (Lead Agent) decomposes tasks, assigns them to Teammates (worker agents), and merges results. Members run in isolated contexts and communicate via an IPC bus.

Differences from Subagents: Subagents act as workers reporting to a single lead, while AgentTeams allow collaborative discussion and parallel execution.

Enabling AgentTeams

Update Claude Code to the latest version.

Install tmux and iterm2 (e.g., brew install tmux).

Enable the experimental flag in ~/.claude/settings.json:

{
  "env": {"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"}
}

Start a new Claude Code session inside tmux (run claude).

Request a team via natural language, e.g.:

# e.g. 1
Create a team to explore CLI tool design: one for UX, one for architecture, one as a challenger.

# e.g. 2
I need a new user authentication module. Build a team: front‑end login page, back‑end Auth API, and test cases. Act as Tech Lead.

Select a teammate display mode: in‑process (all members in the same terminal, switch with Shift+Up/Down) or split‑pane (tmux panes for each member).

Best Practices

Clear CLAUDE.md : A high‑quality CLAUDE.md serves as a shared employee handbook defining coding standards and architecture decisions.

Task Granularity : The Lead Agent’s ability to decompose tasks determines success; overly vague tasks cause workers to stall.

Cost Monitoring : Running multiple parallel Claude instances increases token consumption; reserve AgentTeams for complex refactoring or feature development, and use single‑agent mode for simple bug fixes.