Claude Code 官方教程
图文精读版

00:04 Claude Code is an agentic coding tool that understands your code base, edits your files, runs commands, and integrates with your existing developer tools to help you get things done faster. It's available in your terminal, Visual Studio Code, the Claude desktop app, on the web, and JetBrains IDEs. But we'll be using it in the terminal for this video.

00:41 Unlike Claude AI, Claude Code has direct access to your files in your terminal and your entire code base. So instead of copying and pasting code back and forth, it can go in and do all the work itself. The easiest differentiator is that Claude Code works as an AI agent — software that can interact with its environment and perform actions to complete a defined goal.

01:13 It can read and understand your code base, execute your build script, run your tests, install packages, and use the output to decide what to do next. Claude Code can search the web if it needs documentation on the latest API references.

01:44 First is the context window — Claude's working memory. Second, it asks for permission before running commands or making changes. Third, it can make mistakes — it might misunderstand your intent, introduce a new bug, or over-engineer a solution.

02:17 Quick recap: Claude Code is an agentic coding tool. It reads your code base, edits your files, runs commands, and connects to external tools to help you ship faster.

00:04 If you're on macOS, Linux, or WSL, use this curl command to install it in one go. If you prefer Homebrew, you can use brew install, but note this doesn't have auto-update. For Windows, use the PowerShell Invoke-RestMethod command, a curl command in CMD, or a winget command.

00:35 Go to your project directory and run claude. You go through initial instructions like choosing your theme and signing in with your Claude account (Pro, Max, or Enterprise) or an API key. Whatever directory you run claude in, it has access to that directory and all its subfolders.

01:05 In VS Code, open the extensions panel and search Claude Code — the extension by Anthropic with the blue check. Restart VS Code, then open the command palette with Ctrl/Cmd+Shift+P. You can also opt out of the UI and use the terminal directly in your settings file.

01:34 For JetBrains IDEs, install the Claude Code plugin from the Marketplace and restart. On Claude Desktop, after install and sign-in you'll see a Code toggle at the top — it lets you work in a specific folder, change permissions, even work in a cloud environment.

02:10 On the web, go to claude.ai/code — similar to desktop but restricted to GitHub repositories only. The terminal is the best bet to stay up to date; features ship there fastest. Desktop is great for background work; web is great for remote GitHub work or parallel sessions.

00:04 Claude Code is best explained through the agentic loop. You enter a prompt; Claude Code gathers context, interacts with the model which returns text or a tool call, takes action (editing a file or running a command), then verifies those results.

00:41 If they achieve the goal, Claude finishes and waits; if not, it runs the loop again until results are complete and verifiable. Throughout, you can add context, interrupt, or steer the model.

01:14 Once you reach the limit, Claude Code compacts your conversation. Tools are the backbone of how agents work — they let Claude Code execute code to get closer to a task, e.g. a read-file tool or search-web tool.

01:47 Claude Code has permission modes. Default asks explicit permission. Auto-accept edits files without asking but still asks for commands. Plan mode uses read-only tools to compile a plan. Be cautious when skipping permissions.

02:18 Claude Code combines an agentic loop, a managed context window, tools, and configurable permissions into your terminal — fundamentally different from a chat window.

00:04 You talk to Claude Code like any AI assistant. You can choose whether Claude auto-accepts every file change or requires explicit permission each time. With Shift+Tab you cycle between modes. In auto-accept mode it auto-approves file edits but asks permission to run commands.

00:34 Within the Shift+Tab menu is plan mode. It uses read-only tools to analyze your code base, asks clarifying questions, and returns a long detailed plan. Great for planning complex changes or a safe code review.

01:04 I have an app that needs a dark mode toggle. In the root directory I hit Shift+Tab into plan mode and write the prompt: "My app needs a dark mode implemented across the entire app. Can you create a toggle switch on the header… find a good contrast color based on my existing light theme."

01:32 After reviewing, it looks good. At the end you see explicitly what Claude did and how it reached its conclusion. Try to be as descriptive as possible; use plan mode to get into the nitty-gritty before executing.

00:02 One of the most useful parts is the CLAUDE.md file — persistent memory about your project. Without it, Claude has to start fresh, re-explore the code base, and make assumptions.

00:32 It's a markdown file at the root of your project; Claude Code reads it automatically every session — like an onboarding script. Its contents are appended to your prompt. Run /init to make Claude generate one from your code base.

01:02 Example: a Next.js 15 app using the app router, Tailwind, Drizzle ORM. Commands: dev server, run tests, lint. Code style: two-space indentation, named exports, API routes in app/api, prefer server actions.

01:38 There's a hierarchy: a project-level CLAUDE.md in the root (shared via version control), and a user-level CLAUDE.md in your config folder, just for you across all projects — put personal preferences there.

02:05 If you correct Claude, explicitly ask it to save to memory. Reference docs with the @ symbol and a file path. We recommend starting without a CLAUDE.md so you see where you must course-correct, keeping it compact.

02:34 The difference between a frustrating session and a productive one comes down to context — the CLAUDE.md file is how you provide it. Start with your stack, preferences, then commands.

00:04 If you take one thing away, let it be this workflow: Explore, Plan, Code, Commit. The fastest way to handle steps one and two is plan mode — Claude can't edit files, just reads to gather research. Hit Shift+Tab until you see plan mode.

00:38 Claude reads relevant files, does web searches, gives a plan. Review it; this is the best place to course-correct because it's before any code is written. You can also use explore without plan mode by asking Claude to explore your code base.

01:08 Once the plan looks good, select approve and let Claude toggle through the list items. After the plan is finished we have the context of how it got to the results to guide its next decision.

01:35 For Claude to be confident, it must be clear on what it deems correct — make this explicit. Add tools: for web UIs, the Claude in Chrome extension lets Claude Code control a tab and test the UI. Include a test suite it can validate on.

02:06 Make tests a source of truth to avoid false positives. If Claude keeps running into the same issues, ask it to save the solution to its CLAUDE.md. Before you commit, run a sub-agent code reviewer.

02:35 Then get Claude to generate a commit message in your style. Rinse and repeat. Explore gives context, Plan creates the action plan, Code is the back-and-forth, Commit reviews and pushes your code.

00:00 Context is Claude's working memory. Every file it reads, command it runs, message you send takes up space in the context window. It's finite, so optimizing it is extremely important.

00:39 When you approach the limit, the context window is automatically compacted — it summarizes and removes unnecessary tool call results (may lose details). Run /compact manually to compact up to that point while keeping memory.

01:08 To start completely fresh, run /clear. To check state, run /context — a big picture of your context size and what categories take the most.

01:37 Rule of thumb: working a feature and going over but need to continue → compact. Finished a plan and starting new → clear. For things to remember across sessions, put it in CLAUDE.md.

02:04 The irony of a smaller prompt: long-run it takes more context. Without being explicit, Claude looks around your codebase more and does its own thinking — far more context than a clearer sentence or two.

02:36 MCP servers load all tools into context by default — turn off unrelated ones. Skills work similarly but don't put the whole thing into context. Sub-agents run in parallel with a separate context window; for answer-only tasks, return just a summary.

03:09 Use /compact for long sessions and /clear to start fresh. Be specific. Check what's using context. Use sub-agents to delegate answer-only tasks.

00:03 Model Context Protocol is an open standard that lets Claude Code connect to external tools and data sources. A lot of your context lives elsewhere — databases, productivity apps, public repositories. This is where MCP comes in.

00:37 Tools give agents the ability to perform actions. If your team uses Linear, add a Linear MCP server to bring in issue details. For up-to-date dependency docs, the Context7 MCP server provides that.

01:06 Hundreds of connectors at claude.com/connectors. Add with the claude mcp add command. HTTP servers are for remote services; STDIO servers for local processes. Use /mcp inside a session to see what's connected.

01:40 Scoped three ways: local (current project, you), user (all your projects), project (a .mcp.json checked into version control so the whole team gets the same servers automatically).

02:12 MCP servers add tool definitions to context even when unused. Run /mcp and disable unused ones. If a tool has a CLI equivalent (gh, aws), the CLI is more context-efficient.

02:45 A skill loads just a name and description into context, loaded fully only when needed. If MCP tools exceed 10% of your context window, Claude Code switches to tool search mode. Recap: add with claude mcp add, scope with .mcp.json, watch context by disabling unused servers.

00:03 Hooks let you run commands at different points in Claude Code's life cycle. Unlike everything else, hooks are deterministic — they always run. A CLAUDE.md instruction is followed most of the time but not always; a hook makes it happen every single time.

00:42 Hooks are configured in your settings.json. Pick an event, optionally a matcher for which tool, and a command. Events: UserPromptSubmit, PreToolUse, PostToolUse, Notification, Stop.

01:16 The most common hook: auto-formatting after edits — a PostToolUse hook with a matcher of Edit or MultiEdit. The command checks the file extension and runs the formatter (Prettier for TS, gofmt for Go, Ruff for Python).

01:55 PreToolUse hooks can block tool calls. The hook gets the tool name and input as JSON on stdin. Exit code 0 means proceed; exit code 2 means block, and stderr is fed back to Claude. Block prod config writes, rm -rf, commits to main.

02:31 Hooks in .claude/settings.json are project-level and can be checked into your repo so the whole team gets them. Use the CLAUDE_PROJECT_DIR env var to reference project scripts. If it needs to happen every time without fail, put it in a hook, not a prompt.