Claude Code Complete Guide | 808AI 五金行

1. What is Claude Code?

Claude Code is a CLI AI coding assistant by Anthropic that runs directly in your terminal — reading/writing files, executing commands, managing Git. Not just code completion, but a true AI development agent.

💡 Why Claude Code?

Estimated annual revenue of $2.5 billion, making it the fastest-growing AI development tool. Supports parallel Sub-agents, complete Hooks system, CLAUDE.md persistent memory, and Command → Agent → Skill orchestration workflow.

vs Other Tools

⚡

Claude Code

CLI AI Agent

Full terminal access
Parallel Sub-agents
Complete Hooks system
CLAUDE.md memory
$2.5B revenue (est.)

✈️

GitHub Copilot

IDE Plugin

IDE inline completion
Inline code suggestions
No agent capabilities
No custom Hooks
No memory system

🖱️

Cursor

AI IDE

Graphical IDE interface
Built-in AI chat
Limited agent features
無 Hooks 系統
No native orchestration

2. Core Architecture: The Three Cs

Claude Code's extensibility is built on three components: Agents, Commands, and Skills.

🤖

Agents

16 Fields · 6 Official Built-in

Custom models & tools
Background execution mode
Hooks lifecycle
Git Worktree isolation
Skill preloading

⌨️

Commands

13 Fields · 64 Official Commands

/slash command trigger
Context fork isolation
Model & effort override
Tool permission control
Shell script embedding

🧩

Skills

13 Fields · 5 Official Built-in

Path-matching auto-load
Agent preloading
Independent fork execution
simplify / batch / debug
loop scheduled repeat

🤖 Agents — 16 Frontmatter Fields

Field	Type	Description
name	string	Unique identifier (lowercase + hyphens)
description	string	When to invoke. Use `"PROACTIVELY"` for auto-invocation
tools	string/list	Tool allowlist, supports `Agent(type)` syntax
disallowedTools	string/list	Tools removed from inherited list
model	string	`haiku`、`sonnet`、`opus` 或 `inherit`
permissionMode	string	`default`、`acceptEdits`、`dontAsk`、`bypassPermissions`、`plan`
maxTurns	integer	Maximum agentic turns
skills	list	Skill names preloaded at startup
mcpServers	list	MCP server connections
hooks	object	Lifecycle Hooks
memory	string	Memory scope: `user`、`project`、`local`
background	boolean	Set to `true` for background task execution
effort	string	`low`、`medium`、`high`、`max`
isolation	string	Set `"worktree"` to run in temporary git worktree
initialPrompt	string	First prompt auto-submitted when running as main agent
color	string	CLI output color (e.g. `green`、`magenta`）

6 Official Built-in Agents

#	Agent	Model	Tools	Purpose
1	general-purpose	inherit	All	複雜多步驟任務 — 預Set 代理類型
2	Explore	haiku	Read-only	Fast code search and exploration
3	Plan	inherit	Read-only	Pre-implementation planning
4	Bash	inherit	Bash	Execute terminal commands in separate context
5	statusline-setup	sonnet	Read, Edit	Set 定狀態列
6	claude-code-guide	haiku	Glob, Grep, Read, WebFetch, WebSearch	Answer Claude Code feature questions

⌨️ Commands — 13 Frontmatter Fields

Field	Type	Description
name	string	Display name and `/slash-command` ID
description	string	Description shown in autocomplete
argument-hint	string	Argument hint for autocomplete
disable-model-invocation	boolean	Prevent Claude from auto-invoking this command
user-invocable	boolean	Set `false` to hide from menu
paths	string/list	Glob patterns to limit loading
allowed-tools	string	Tools allowed without permission prompts
model	string	Model used when running
effort	string	Override effort level
context	string	Set `fork` to run in isolated subagent
agent	string	Subagent type when context: fork
shell	string	`bash`（預Set ）或 `powershell`
hooks	object	Lifecycle hooks scoped to this command

64 Official Commands (by Category)

#	Command	Category	Description
1-3	/login /logout /upgrade	Auth	Account login, logout, upgrade plan
4-15	/color /config /keybindings /permissions /privacy-settings /sandbox /statusline /stickers /terminal-setup /theme /vim /voice	Config	主題、按鍵、權限、沙盒、語音等Set 定
16-22	/context /cost /extra-usage /insights /stats /status /usage	Context	Context visualization, token usage, analytics
23-27	/doctor /feedback /help /release-notes /tasks	Debug	Diagnostics, bug reports, background tasks
28-29	/copy /export	Export	Copy response to clipboard, export conversation
30-37	/agents /chrome /hooks /ide /mcp /plugin /reload-plugins /skills	Extensions	Agent, MCP, Plugin, Hooks management
38	/memory	Memory	編輯 CLAUDE.md memory檔、管理 auto-memory
39-43	/effort /fast /model /passes /plan	Model	Model switching, effort adjustment, Plan mode
44-49	/add-dir /diff /init /pr-comments /review /security-review	Project	Working directory, Diff viewer, PR, security review
50-56	/desktop /install-github-app /install-slack-app /mobile /remote-control /remote-env /schedule	Remote	Desktop App, Slack, GitHub, remote control, cloud scheduling
57-64	/branch /btw /clear /compact /exit /rename /resume /rewind	Session	Branch, compact, rewind, Session management

🧩 Skills — 13 Frontmatter Fields

Field	Type	Description
name	string	Display name and `/slash-command` ID
description	string	Description for auto-discovery
argument-hint	string	Argument hint for autocomplete
disable-model-invocation	boolean	Prevent Claude from auto-invoking
user-invocable	boolean	Set `false` to hide (agent preload only)
allowed-tools	string	Tools allowed without permission prompts
model	string	Model used when running
effort	string	Override effort level
context	string	Set `fork` to run in isolated subagent
agent	string	Subagent type when context: fork
hooks	object	Lifecycle hooks scoped to this skill
paths	string/list	Glob patterns to limit auto-loading
shell	string	`bash`（預Set ）或 `powershell`

5 Official Built-in Skills

#	Skill	Purpose
1	simplify	Review changed code — refactor to eliminate duplication, improve quality
2	batch	Run commands across multiple files in bulk
3	debug	Debug failing commands or code issues
4	loop	Run prompts on a recurring schedule (up to 3 days)
5	claude-api	Build apps with Claude API / Anthropic SDK

3. Memory System

CLAUDE.md is Claude Code's persistent memory — loaded on every startup, letting Claude remember your project conventions, preferences, and workflows.

📝

Writing a Good CLAUDE.md

CLAUDE.md is the most impactful way to improve Claude Code output:

Project structure
Code style conventions
Common commands & paths
Test & deploy workflows
Personal preferences via CLAUDE.local.md

🗂️

Three-Layer Loading

Claude Code loads memory in two directions:

Ancestor loading (UP) — walks up from cwd at startup, loads all CLAUDE.md files
Descendant loading (DOWN) — subdirectory files only load when accessed
Global — ~/.claude/CLAUDE.md applies to all sessions
Sibling directoriesnever load

Monorepo Structure Example

/mymonorepo/
├── CLAUDE.md ← 根層共用（啟動即 load）
├── frontend/
│   └── CLAUDE.md ← before端專屬（讀取時才 load）
├── backend/
│   └── CLAUDE.md ← after端專屬（讀取時才 load）
└── api/
    └── CLAUDE.md ← API 專屬（讀取時才 load）

💡 load範例

從 /mymonorepo/frontend/ Starting from here:
✅ /mymonorepo/CLAUDE.md — 祖先目錄，啟動即 load
✅ frontend/CLAUDE.md — 當before目錄，啟動即 load
❌ backend/CLAUDE.md — 不同分支，不 load
❌ api/CLAUDE.md — 不同分支，不 load

4. Hooks: Deterministic Guardrails

Hooks are deterministic scripts that run outside the agentic loop — triggered on specific events, providing safety guardrails and automation.

⚠️ Core Principle

Hooks trigger after tool operationscompletetriggers (PostToolUse), not before.PreToolUse is the exception — it can intercept before execution.

6 Hook Types Supported by Agents

🔍

PreToolUse

Triggersbeforetool execution. Can block, modify, or log pending operations.

✅

PostToolUse

Triggersaftertool execution. Validate output, run linters, send notifications.

🔔

Notification

Triggers when Claude needs user attention. Connect to sounds, desktop notifications.

🛑

Stop

Triggers when agent completes a turn. Run final checks or cleanup.

🔐

SubagentStop

Triggers when subagent stops. Collect child agent results.

🌐

HTTP Hooks

Trigger external webhooks via HTTP POST with JSON payload.

Set 定範例

settings.json

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "npx eslint --fix $CLAUDE_FILE_PATHS"
          }
        ]
      }
    ],
    "Notification": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "afplay /System/Library/Sounds/Glass.aiff"
          }
        ]
      }
    ]
  }
}

5. Orchestration Workflow

Command → Agent → Skill is Claude Code's core orchestration pattern. A Command starts the flow, Agent handles execution, Skill provides domain knowledge.

⌨️

Command

Entry point · User interaction

→

🤖

Agent

Execute tasks · Preloaded Skill

→

🧩

Skill

Domain knowledge · Output results

Example: Weather System

Component	Role	Example
Command	Entry point, user interaction	/weather-orchestrator
Agent	Fetch data with preloaded skill	weather-agent(preloaded weather-fetcher)
Skill	Independent visual output	weather-svg-creator

Execution Flow

// User input
/weather-orchestrator

// Step 1: Command asks user preference
├─ AskUser: "Celsius or Fahrenheit?"
│  └─ User: "Celsius"

// Step 2: Command invokes agent via Agent tool
├─ Agent tool → weather-agent
│  ├─ Preloaded Skill: weather-fetcher(domain knowledge)
│  ├─ Fetches from Open-Meteo API → 26°C
│  └─ Returns temperature to Command

// Step 3: Command invokes skill via Skill tool
├─ Skill tool → weather-svg-creator
│  ├─ Output → weather.svg(weather card)
│  └─ Output → output.md(summary)

// Done
└─ Display results to user

💡 Two Skill Usage Patterns

Agent Skill（預載）: Injected via agent's skills field as domain knowledge.
Skill（獨立）: Invoked independently via Skill tool, produces output directly.

6. Practical Tips

Practical tips for everyday Claude Code usage to boost your productivity.

📊

/context — Context Visualization

查看當before context 使用量的彩色格狀圖。顯示哪些工具佔用最多空間，建議如何優化。

🔀

context: fork

在 command 或 skill 中Set context: fork，讓它 to run in isolated subagent，不污染主對話。

📂

paths Pattern Matching

用 paths 欄位Set 定 glob 模式，讓 skill 只在操作特定檔案時自動 load。

🤖

Auto Mode

用 claude --enable-auto-mode to start. Background safety classifier replaces manual permission prompts.

⚡

effort Adjustment

/effort low|medium|high|max to adjust thinking depth.max requires Opus 4.6.

🔄

/compact — Compress Conversation

When context is nearly full, use /compact to compress. Add focus:/compact keep API content

Common CLI Startup Flags

Flag	Short	Description
--continue	-c	Continue most recent conversation
--resume	-r	Resume specific session
--model <NAME>		Specify model (sonnet, opus, haiku)
--print	-p	Non-interactive output (headless / SDK)
--permission-mode		default、plan、acceptEdits、bypassPermissions
--allowedTools		Tools allowed without permission prompts清單
--add-dir <path>		Add working directory to session
--worktree		Run in git worktree
--enable-auto-mode		Enable Auto Mode (beta)

7. Installation & Getting Started

Three steps to start using Claude Code.

Terminal

# 1. Install Claude Code
npm install -g @anthropic-ai/claude-code

# 2. Navigate to your project
cd your-project

# 3. Start Claude Code
claude

Quick Reference

🚀 Basic Operations

claude                # Start interactive mode
claude -c             # Continue last conversation
claude -r             # Resume specific session
claude -p "問題"     # One-shot Q&A
claude /init          # Initialize CLAUDE.md

⚙️ Advanced Operations

claude --model opus   # Specify model
claude --worktree     # Git worktree isolation
claude --enable-auto-mode  # Auto Mode
/context               # Check context usage
/compact               # Compress conversation
/effort max            # Maximum thinking depth

💡 808AI Recommendation

First time? Run claude /init to create CLAUDE.md and write your project conventions in it.
This is the highest ROI step — Claude remembers your preferences, making every conversation more precise.