Last updated on

Demystify AI by Using It

AI will be a nightmare for those who refuse to think harder. Innovation of machines won’t liberate people from work; it aggravates it. History speaks loudly. 🥸

The idea to write this article comes from a post by Andrej Karpathy on X.

Why Claude Code

Skip this if you don’t like listening to unprofessional long-winded speeches and blablabla… It’s totally OK! HAHAHA 🤪

Claude Code is not the only option. There are a lot of AI Agents actually, and impressively, most of them are open-source. For example, Openclaw, Hermes agent, Deepseek TUI. However I don’t recommend them to a novice, especially when you want to use them as productivity tools to build something for work that really matters. Because it’s energy-draining to maintain a machine from scratch with precision and effectiveness. Don’t waste your time.

Claude Code CLI stands out. To be exact, in this article we talk about Claude Code CLI (CCL) rather than the one that lives in the desktop app with UI — the former basically runs in the Terminal. Actually the Co-work and Claude Code features (if you know) in the desktop app are derivatives of CCL. I have to admit that the desktop app is more novice-friendly. However, the installation and use of the desktop app relies on a Claude AI subscription and “special” network environments — both unreachable for many — and the same goes for CodeX from OpenAI. Anyway, if these issues are not a problem for you, then just use the desktop apps, which are definitely easier.

What and how you can do. Back to the subject, CCL is one of the most powerful tools to date that links user, computer and LLM (beyond the conventional chatbot — like chatting with DeepSeek on the web — which only links user and LLM). It can execute commands, write scripts, generate files, analyze, debug… almost everything a computer can do, even destroy itself and your Operating System (OS), through plain conversation. But don’t worry — one reason we choose CCL over the open-source alternatives is safety (under reasonable supervision by users, of course). As standalone software, CCL can be deployed locally with a few commands in a clean network environment on any OS, totally free and accessible. With the help of some amazing software (CC switch, will mention later), CCL can connect to any available LLM API, like DeepSeek, DOUBAO and Qwen. CCL ➕ CC switch ➕ customized LLM API — with a few commands and clicks (plus payment for the API) — you get a capable AI Agent that can do many things that previously could only be done by people with professional CS backgrounds. Mostly, remember: you achieve this through conversation alone.

The only catch. The biggest friction comes from the Terminal. Non-CS-professionals will naturally fear the Terminal or anything without graphical windows. But remember, all you need to do in that terminal is type words and have conversations — just like on a regular chat page.

TL;DR: Claude Code CLI is Production-ready, High-performance and Industry-leading. Most importantly, it’s reachable for any user. The only trade-off is that one has to learn (a little effort) and then accommodate to the Terminal.

Before You Start

Wrote this after my lovely friend cleaned his whole /home directory by mistake under my supervision. Disaster. 😥😕🤓

  • Highly recommend to backup the important files/directories/system (best for regularly) to minimize your losses in an accidental crash, which is quite normal. Try Time machine if you’re macOS user.
  • Better Copy/Paste the command than Type On Your Own. (If you can’t figure out the difference between rm -rf ~/.claude and rm -rf ~/ .claude) There is a “copy” button at right of each command in this article, click it!
  • Follow the instructions IN Order to prevent unnecessary trouble.
  • Feedback and help-asking are welcomed.

Installation

Claude Code CLI step-by-step

Official Claude Code Docs here.

Windows: Open the Terminal first. Right click the WIN button (Start button) and choose "open the terminal (administrator)".

Once you open the Terminal, copy this command to that window and press Enter:

winget install Anthropic.ClaudeCode --source winget

Tips: Installation of Git for Windows is officially recommended — “so Claude Code can use the Bash tool” — which is quite easy, and just treat it as a very useful plugin for CCL itself. Git for Win is optional, not necessary. (BTW, you can also get a “Git Bash” app that allows users to run a Bash shell on Windows, but I guess you’ll use WSL if you truly know about Bash anyway.🥱)

macOS: Recommend to use Homebrew to install and manage softwares on macOS.

  • Install Xcode first:
xcode-select --install
  • Run this command to install Homebrew:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
  • Try installing CCL via Homebrew directly:
brew install --cask claude-code

If the process is stuck for a long time (>1 min), press ctrl+C to end it and try the following way.

  • Install npm and configure domestic sources, in order:
brew install npm
brew install nrm
nrm use taobao
  • Install CCL via npm:
npm install -g @anthropic-ai/claude-code

Note: Installation via npm is deprecated officially. Which means that you may not get subsequent updates in future. However, it’s the only way to get this software quickly and easily in a domestic network environment. But don’t worry, I’m sure there will be more reachable and powerful alternatives soon.

Linux/WSL/macOS: I do believe that Linux users all left at the very beginning of this article (lol). You can figure it out on your own 🤓. Anyway:

curl -fsSL https://claude.ai/install.sh | bash

CC switch

CC switch is an app designed to switch the LLM model setting of AI agent software (such as Claude Code, CodeX, Openclaw, …) to any customized model provider. It runs with a UI — you just need to fill in your API key in the app, press the Save button and Enable that option in the list. Then it will automatically find your agent’s path and change the setting, effective immediately.

An open-source software, totally free. Download from here. Find the version suitable for your OS under Assets. Then install — easy.

API

Get the API key from model providers according to your personal preferences and budget. Some links: DeepSeek, Kimi, Minimax, SiliconFlow…

Note: Steps for Windows and macOS are verified by the author, OS versions are Windows 10 & macOS 26. No errors. The Linux step is not verified.

Getting started

  • Close the Terminal to make all the changes take effect.
  • Decide on a working directory that stores any necessary files you want the AI to read/write or generate in the future.
  • Navigate to that directory in the Terminal, or open the Terminal under that directory directly. Here about how to open Terminal in Finder, macOS. And For Windows10 and similar for Windows11 without pressing shift.
  • Type claude and press Enter. Bingo! 👾
  • On first launch, you’ll need to configure some settings by following the guidance. Choose using ↑↓ and Enter — easy.
  • Enjoy the magic! 😺

Tips

  • Type / — the bottom bar shows all available functions in CCL in a [/name] instructions format. Select with ↑↓ and Enter, then Enter again to see what happens.
  • Use /powerup to quickly discover what CCL can do.
  • There are no format limitations for prompts — just have a conversation. Talk to the LLM about whatever you want.
  • Let the agent solve anything you encounter on your computer, even its own settings or errors. YES! The agent can even modify itself.
  • CCL will frequently ask for your permission when doing things it considers dangerous — think before you authorize. Generally, CCL’s permissions to access files/settings are limited to the working directory where you launched CCL (via the claude command). It’s relatively safe, but consider an extra backup before letting the AI change anything. You can ask the AI to do the backup itself — just tell it “backup before changing (something like that).”
  • Watch the token budget! Check your bill regularly.

Future outlook

  • Working principles of modern AI agents
  • Key features of Claude Code
  • More tips on collaborating with Claude Code
  • Flibbertigibbeting…

Bottom lines

I think everyone should engage with AI firsthand to explore its true capabilities and limitations. By iteratively adapting our workflows and thinking styles to this collaboration, we must recognize that AI is not merely a tool; it is a powerful force actively reshaping human consciousness and even the physical structure of the brain itself.