60分钟全面掌握Claude Code (Master Claude Code in 60 Minutes)

Creator: 秋芝2046 | Duration: 56 min | Date: 2026-05-05

Part 1: Getting Started (Basics)

What is Claude Code?

• Created by Anthropic in Feb 2025; a terminal-based AI Agent (nicknamed "CC")
• Goes beyond Q&A — it plans and takes actions autonomously using the LLM Loop: prompt → LLM thinks → calls tools → gets results → repeats until done
• Key advantages over other agents:
  1. Runs locally — direct read/write access to your files, terminal commands
  2. Superior Harness engineering — the non-LLM design around the model that dramatically improves performance (same model, different harness = different results)

Use Cases

• App development & coding (its origin)
• Copywriting, research, info organization
• Data analysis, spreadsheets, reports
• Content creation, script analysis, asset matching

Installation

• 4 ways to run CC: desktop app, web browser, IDE plugin, terminal (recommended — most native, latest features)
• Pair with an IDE (Cursor, Trae, etc.) for file browsing + manual editing + terminal access
• Method 1: One-line install from Claude Code website → paste in terminal
• Method 2: Have the IDE agent install it for you (handles dependencies/network issues)
• Verify: claude --version

LLM Configuration

• Best: Claude subscription → login command
• Alternative: Domestic/third-party APIs via CC Switch tool — configure API key + base URL before launching CC
• CC Switch also lets you assign models to high/medium/low tiers

First Launch

• Type claude → initial setup (language, theme, safety prompts, trust folder)
• Ready to chat

Part 2: Interaction & Permission Modes

Three Permission Modes (cycle with Shift+Tab)

• A 4th "YOLO" mode exists (added at launch command) — no confirmations at all. Anthropic found 93% of users just approve anyway.

Ways to Interact

1. Text conversation — basic prompting
2. @ mention files — precise context injection, saves tokens
3. Images — drag or Ctrl+V (even on Mac, not Cmd+V) for visual references
4. Slash commands — /help, /model, /btw (side question outside project context), /simplify (3-agent code review)

Pro Tips

• Shorter prompts = more tokens spent (CC explores more to compensate)
• For long/specific prompts: write requirements in a local file → @ mention it
• Line break in CC terminal: Option+Enter (Mac) / Ctrl+Enter (Windows) — NOT Shift+Enter (that sends)

Part 3: Managing CC (Safety & Context)

Rollback & Version Control

• Quick undo: Esc twice or /rewind — rolls back conversation + file edits (but not terminal commands like installs)
• Real safety net: Git — treat it as a save system
  • CC can install Git, link GitHub, commit, push, rollback — all via natural language
  • Recommended workflow: commit after each successful step

Context Window Management

• LLMs degrade as context fills up (60-80% of advertised window is truly effective)
• /compact — compresses past conversation, keeps key info, frees space
• /clear — wipes everything (or just open a new terminal)
• /context — shows detailed token usage breakdown
• Enable persistent context display via a config command (requires terminal restart)
• Personal rule: compact when usage exceeds ~60%

Resuming Conversations

• /resume — pick from conversation history
• claude -c — continue last conversation directly

Part 4: Personalization (Making CC Work for You)

Layer 1: CLAUDE.md (Highest Priority — Always Loaded)

Three levels:

• Create project-level: /init (analyzes project, writes structure)
• Create/edit global: /memory → select global CLAUDE.md
• Best practice: Start with high-level principles; gradually add lessons from CC's mistakes
• Don't make it too dense — keep it focused on unchanging rules + error corrections

Layer 2: Auto Memory (Second Priority — On-Demand)

• Enable via /memory → toggle on
• A background agent silently records:
  • User info (who you are, role)
  • Feedback (corrections: "not this, do that")
  • Project info (progress, decisions, tech choices)
  • External resources (document locations, references)
• Stored as files in .claude/projects/<path>/memory/
• Only MEMORY.md (index) is loaded initially; details read on-demand
• Project-scoped (doesn't transfer between projects)
• You can tell CC to forget things it remembered incorrectly

Layer 3: Custom Docs (Self-Built Memory)

• Write your own reference files (brand guidelines, style guides, etc.)
• Reference them in CLAUDE.md: "When modifying frontend visuals, refer to brand-guide.md"
• Essentially: "inject compressed context at the right moments"

Part 5: Advanced Extensions

Skills (Professional Instruction Manuals for AI)

• Four types: knowledge-based, process-based, tool-based, hybrid
• Only metadata (name + trigger description) is loaded into context; full content loaded only when invoked
• Install: Place skill folder in .claude/skills/ (global) or project .claude/skills/
• Invoke: Automatically (LLM decides), manually (/skill-name), or explicitly in prompt
• Find skills: Install "Find Skill" skill → search community skills
• Create skills: Install "Skill Creator" skill (by Anthropic) → conversational creation

MCP (Model Context Protocol)

• Adapter connecting AI to external tools/services (Figma, Notion, databases, etc.)
• Downside: token-heavy, hard to run many simultaneously
• Trend: lightweight MCPs → becoming Skills; heavyweight ones → becoming CLIs

CLI (Command Line Interface Tools)

• External services packaged as terminal commands for agent use
• Examples: Feishu CLI (docs, calendar, email), OpenCLI (social media APIs)
• Install: paste tool URL into CC, it handles setup
• More efficient than agents simulating human browsing (screenshots, clicking)

SubAgents (Parallel Workers)

• Each SubAgent has its own context space — doesn't pollute main agent's context
• Can run in parallel for speed
• Auto spawn: CC detects parallelizable work and creates subs automatically
• Manual creation: /agent → guided conversational setup
• Use case example: Main agent codes while SubAgents research competitor apps simultaneously

Hooks (Automated Triggers)

• "When CC does X, automatically do Y"
• Examples: play sound on task completion, run code formatter before commits, send Feishu notification
• Configure by telling CC what you want in natural language

Plugins (Bundles)

• A package of Skills + SubAgents + Hooks + MCPs combined
• Manage via /plugin → discover and install from marketplace

Summary: The Four Stages

Core insight: You're learning to move from Q&A with AI to systematizing its autonomous work — a transferable skill for any future AI tools.