Claude Code Getting Started: Installation, Switching to Chinese LLMs, and Git Workflow Guide

The Evolution of Claude Code and AI Programming Tools

Claude Code (referred to as CC throughout this article) has rapidly become one of the most talked-about AI programming tools, evolving from a simple command-line code assistant into a genuine "engineering collaboration partner." CC is developed by Anthropic — an AI safety company founded by former OpenAI researchers, known for its "Constitutional AI" training methodology and emphasis on model interpretability. The Claude family of large language models that powers CC has notable strengths in long-context window processing and code reasoning, which is the technical foundation enabling its "global context awareness." This article, compiled from a hands-on tutorial by a Bilibili content creator, systematically covers CC's core capabilities, environment setup, switching to Chinese domestic models, and a complete Git workflow implementation — helping developers get up to speed quickly and avoid common pitfalls.

The Evolution of Claude Code's Core Capabilities

Claude Code started with a command-line interface and relatively limited functionality: analyzing project code, understanding relationships between files, and offering modification suggestions. With rapid iteration, today's CC has undergone a fundamental transformation.

From single-file understanding to global context awareness. Where it once analyzed only individual files, CC can now answer detailed questions about an entire project's structure with impressive depth. This leap in contextual awareness is powered by the Claude model family's extremely long context windows (up to hundreds of thousands of tokens), allowing the model to simultaneously "read" the relationships between hundreds of files in a large codebase rather than processing them in isolation — a key shift from being a "tool" to becoming a true "collaborator."

From coding assistant to full engineering lifecycle coverage. CC can now not only write code, but also produce SDD (Specification-Driven Development) documents, write automated test logic, handle deployments, and even generate video content. SDD (Specification-Driven Development) is a development methodology centered on formal specification: developers first write detailed functional specification documents describing what the system should do (rather than how to implement it), then generate code and test cases based on those specs. CC's ability to produce SDD documents means it can transform natural-language business requirements into structured specifications that guide subsequent code generation, significantly reducing requirement misinterpretation. Developers can also use the Skill mechanism to codify real-world working experience into reusable skills, dramatically boosting efficiency.

CC delivers distinct value depending on your role:

• Beginner developers: Helps read hard-to-understand open-source project code, locate legacy bugs and performance bottlenecks, and fix them;
• Solo developers: Through Multi-Agent Orchestration, simultaneously play the roles of product manager, developer, and tester, enabling integrated full-stack development. Multi-Agent Orchestration is an architectural pattern where complex tasks are broken down and completed by multiple AI agents working in parallel or in sequence — a primary agent handles task planning and scheduling while sub-agents handle specialized tasks, letting a solo developer simulate the collaborative efficiency of a full team at minimal cost;
• Team collaboration: Use CLAUDE.md to standardize development conventions, build custom Skills for code review and issue analysis, lower communication overhead, and help new team members ramp up quickly. CLAUDE.md is similar to .editorconfig or .eslintrc — a special configuration file automatically recognized by CC where teams can define code style, architectural conventions, and prohibited operations, ensuring all AI-generated code conforms to established standards and fundamentally solving the consistency problem of "every AI doing its own thing."

In short, humans focus on setting goals, applying constraints, making judgments, and controlling quality — while execution, analysis, comparison, and repetitive work are delegated to CC.

Environment Setup and Switching to Chinese Domestic Models

CC currently comes in three versions: a Web version (use directly in the browser), a CLI command-line version, and a plugin version integrated into VS Code / PyCharm / IDEA. For day-to-day development, the command-line and editor plugin versions see the most use.

CLI Installation Steps

Installing the CLI version is straightforward — a single command is all it takes. Choose the appropriate command for your OS: Mac, Linux, and WSL use the official script, while Windows has its own install command. After installation, run claude -v to check the version number — if you see version info, installation was successful.

Prerequisite: Install a Node.js environment first. Download and install it from the Node.js official website (the tutorial demonstrates version 22.2.0), then run node -v to verify. If you encounter network issues during npm installation, switch to a Chinese mirror source to speed things up.

Updating is equally simple — just run claude update. CC will also proactively prompt you to upgrade when it detects a new version.

Switching to Chinese Domestic LLMs with CC Switch

Due to network access restrictions on the official models, users in China are advised to use the open-source tool CC Switch to switch to domestic large language models for a more stable experience. CC Switch works as a local proxy, intercepting requests Claude Code sends to the Anthropic official API and forwarding them to domestic model endpoints that are compatible with the OpenAI API format. Since mainstream Chinese models like DeepSeek, Kimi, and Qwen all provide OpenAI API-compatible interfaces, CC Switch only needs to handle protocol adaptation and format conversion to enable seamless switching.

Steps to switch:

1. Download the appropriate version from CC Switch's release page. Windows users are recommended to use the portable version (just unzip and run). Note the difference between ARM and AMD architectures — ARM applies to Apple M-series chips or some Windows on ARM devices, while AMD64 applies to traditional x86-64 processors. Choosing the wrong architecture will prevent the program from running; AMD64 is usually the right choice. An .msi installer version is also available;
2. Obtain an API key for a domestic model — options include DeepSeek V3, Kimi, Qwen, and others;
3. In the CC Switch interface, click "+" to add the model, enter the API key (keep the endpoint address at its default), then click enable;
4. Go to Settings and check "Local Routing," then restart CC for the changes to take effect;
5. While using CC, run the slash command /model to view and switch to DeepSeek or other models.

The experience with domestic models is broadly comparable to the official model, though there may be minor differences in output quality. If you have an official paid subscription, you can log in directly through the official authentication flow.

Getting Started: Quick Tips and Common Commands

The first thing most new users do with CC is ask it to analyze a project's structure. Whether you're a new intern or encountering an unfamiliar codebase for the first time, simply type "help me analyze this project" and CC will automatically identify the project type (e.g., a Vue-based frontend admin system) and walk through the architecture module by module.

For users unfamiliar with CC's commands, the easiest approach is to ask directly in conversation — for example, "which command should I use to resume my last conversation?" CC will tell you exactly what to do, no need to dig through official documentation.

Commands fall into two categories — internal slash commands and external commands:

• claude -c / --continue: Resume the last conversation (used frequently);
• claude -r / --resume: View the conversation history list and choose one to resume;
• claude -p: Headless mode — executes a single task as a command, such as "analyze the Test directory project structure" and outputs results to the console;
• /clear: Clear the current conversation and start a fresh session;
• /help: Look up all available commands.

CC's Code Modification Workflow

Every time CC modifies code, it follows a fixed process: automatically locate the target file → display the specific changes → wait for user approval → execute the modification. This design reflects the important Human-in-the-Loop (HITL) principle in AI systems — requiring human approval before the AI performs critical operations. HITL originated in automation and control systems, and was later adopted in AI ethics and safety research. Its purpose is to prevent destructive changes caused by model hallucinations or misunderstood requirements. Users can approve or reject changes one by one, a design that improves efficiency while preserving human final control over code quality.

Conversational Git Workflow in Practice

One of CC's most practical capabilities is making Git operations feel as natural as everyday conversation. Git is the dominant distributed version control system today, but its command set is complex and has a steep learning curve — especially unfriendly for non-specialist developers. CC's "conversational Git" is powered under the hood by Tool Use / Function Calling: the model understands natural-language intent, maps it to specific Git command sequences, and executes them via the shell. Users don't need to memorize complex Git commands — just describe what you need, and CC automatically calls the right commands.

The tutorial walks through a complete end-to-end example:

1. Create a project: "Help me create a personal tech blog using HTML + CSS + JS";
2. Clone a remote repository: "Help me pull the XX project," and CC automatically runs git clone;
3. Commit and push: "Move the files to the Test project, commit and push the code," and CC automatically handles add, commit, and push, generating a readable commit message based on semantic understanding of the code diff;
4. Query commit history: "Help me check how many commits the Test project has and show them in a table";
5. Branch development: "Create a dev branch, implement three sorting algorithms in it, then commit and push."

Throughout this entire process, the user never manually changed a single line of code — everything was done through conversation. Notably, after writing the code, CC automatically runs tests (all 12 test cases passed in the tutorial) and only commits after confirming everything is correct, ensuring code quality.

Automated Bug Discovery and Fix Loop

CC handles bugs in two modes: one where the user has already found a problem and asks CC to locate and fix it; and one where CC proactively scans the entire project for potential issues. The tutorial demonstrates the latter — asking CC to analyze the dev branch code for potential problems and output the results to an issues.md file.

The identified issues span multiple dimensions: correctness problems, discrepancies between documentation and implementation, design and portability issues, and more. Then, a single instruction — "fix all the issues referenced in issues.md, commit and push the code" — completes the full automated loop from problem discovery to resolution.

The underlying processing flow is: locate relevant code → understand context and root cause → propose and display a solution → run tests to verify the fix. This closed loop essentially chains together static code analysis, semantic understanding, and automated testing into the smallest unit of continuous quality assurance.

Summary

Claude Code has evolved from a code completion tool into a full engineering collaboration platform covering development, testing, deployment, and code review. For developers in China, mastering three key points — CLI installation, switching to domestic models like DeepSeek via CC Switch, and getting comfortable with conversational Git workflows — is enough to integrate CC into your daily development routine quickly. The real productivity gains come from letting AI handle repetitive work while humans stay focused on setting goals and maintaining quality control.

This article is compiled from hands-on tutorial content on Bilibili, for learning and reference purposes only. For actual implementation, refer to the official documentation.

Key Takeaways