大部分开发者第一次接触 Claude Code 时，都会犯同一个错误：把它当成一个更聪明的 ChatGPT。于是对话变成这样——“帮我写个登录接口”，AI 吐出一段代码，复制粘贴，报错，再问，再错。三轮之后得出结论：这玩意儿也就那样。
问题不在模型，在于心智模型错位。传统 IDE 补全式 AI 是"你写它猜"，Agent 式 AI 是"你说它干"——它会自己读文件、跑命令、改代码、看报错、再改。这套能力一旦用对，你实际上是在带一个不知疲倦但需要明确指令的实习生；用错了，你就是在跟一个失忆的聊天机器人扯皮。
claude-code-for-beginners 要解决的就是这个认知落差。它不教你 API 怎么调，而是教你如何组织上下文、如何拆分任务、如何给 AI 设定边界——这些恰恰是官方文档里最容易被一句"just ask Claude"带过、却最难自学的部分。
核心技术揭秘
要理解这套课程的设计逻辑，得先看清 Claude Code 本身的运行闭环。它本质上是一个 读取—规划—执行—回灌 的 Agent 循环，而不是一次性的代码生成：
graph TD
A[用户自然语言指令] --> B[上下文装配<br/>CLAUDE.md + 仓库文件]
B --> C{"任务复杂度判断"}
C -->|"单点改动"| D[直接工具调用]
C -->|"多步任务"| E[拆解为子任务]
E --> F[规划并与用户确认]
F --> G[工具执行<br/>读文件/改代码/跑命令]
G --> H[结果回灌上下文]
H --> I{"验证通过?"}
I -->|"否"| E
I -->|"是"| J[产出变更与总结]
D --> J
这个循环里藏着三个真正的技术难点，也是课程必须覆盖的内容：
其一，上下文装配。CLAUDE.md 不是装饰品，它是这个循环的"系统提示层"，决定了 AI 的项目认知基线。写得好的 CLAUDE.md 会明确技术栈、目录约定、命令入口、禁区（比如"不要动 migrations 目录"）；写得差的，等于让实习生第一天上班没有任何交接文档。课程若能把这一层讲透，价值就已经超过大半同类教程。
其二，任务的粒度控制。 Agent 循环的失败模式几乎都是"任务太大导致上下文爆炸"或"验证信号缺失导致自我循环"。课程需要教会学习者把"重构整个模块"拆成"先改接口签名、再改调用方、最后跑测试"——这本质上是把软件工程里的增量交付思想，翻译成 AI 能消化的指令序列。
其三，权限与副作用边界。 Bash 工具意味着 AI 能真的执行命令。仓库读写、git 操作、依赖安装，每一步都有副作用。课程如果把权限模式（只读 / 逐次确认 / 全自动）当作核心章节，说明作者理解了工程落地的真实风险，而不是只停留在"哇它能自己改代码"。
You installed Claude Code. You typed a few prompts. It wrote some code, broke something else, asked for permissions you didn’t understand, and now you’re not sure if you’re using it right.
You’re not. Almost nobody is at first. I wasn’t either.
The problem
Claude Code is the most powerful coding tool I’ve used, and most people use maybe a tenth of it. They chat with it like a chatbot and never touch the parts that actually change how fast you ship.
The docs tell you what each feature does, not how to use them together. You know plan mode, CLAUDE.md, subagents and hooks exist. Nobody tells you which one to learn first, or how to combine them into a workflow that saves you hours every day.
It changes every week. New commands, new permission modes, new agent features. Half the tutorials you find are already out of date, and you can’t tell which half.
Learning by trial and error is expensive. Every wrong habit costs you tokens, broken code and evenings you don’t get back. Some of those habits (letting it write before it reads, skipping the plan, never reviewing the diff) quietly cost you more than you’d guess.
Meanwhile the people who did figure it out are shipping in an afternoon what used to take them a week.
That gap doesn’t stay the same size. It compounds. Someone who started six months ago isn’t six months ahead of you - they’re building things you haven’t learned to ask for yet.
Why I made this
I use Claude Code every day to build and ship my own products - gitscroll.dev, aicofounders.co, moltplace.net, a Rust MCP framework 🦀 and a few more. I do it around a 9-5 and a kid, so I don’t have time to waste on bad habits, and neither do you.
I learned all of it the slow way. This is the course I wish someone had handed me on day one: the right things, in the right order, kept up to date.
How this course fixes it
Figuring it out alone
This course
Where to start
Docs, random videos, X threads
Module 1, 20 minutes, you’re running
What order
Whatever you stumble on
15 modules, each one building on the last
Practice
Hope you remember it
Hands-on challenges at four levels in every module, with worked solutions
Staying current
Tutorials that broke two releases ago
Updated when Claude Code changes - last update October 2026
Cost
Your evenings and your tokens
Free, no signup, plain Markdown right here on GitHub
You don’t need to be early to the whole AI wave. You just can’t afford to be late to this one. Start Module 1 now →
What’s new
Late September 2026
Plugins & marketplaces (Module 12) - install someone else’s whole setup in one command, and how to vet a plugin before you trust it.
Dynamic workflows (Module 6) - fan a big job out to dozens of agents with /deep-research, “use a workflow”, and ultracode.
Artifacts (Module 15) - publish a live page from your session, no deploy needed.
Auto mode is now the default - permission modes updated across the course, plus the new commands (/plan, /export, /skills, /doctor prompt-audit, /workflows).
Claude + other models (Module 14) - and a full Advanced Module 28 on deterministic gates, routers, and decision models like Jev.
A command-line tool from Anthropic. You describe what you want, and it reads your files, writes and edits code, runs commands, and tells you what it’s doing along the way.
You’ll need a Claude Pro/Max subscription or an Anthropic API key. On first launch it opens a browser to log in, or you can set ANTHROPIC_API_KEY. If anything goes wrong, Module 1 walks through installation on every platform.
Who it’s for
Complete beginners - you don’t need to know how to code, just a computer and a Claude account. Students who want to build portfolio projects with an AI tutor alongside. And experienced developers who want to get faster: skip straight to Module 6 (agents) or Module 12 (skills, hooks, plugins).
It doesn’t care what language you write:
Course map
15 modules, about 9 hours in total. Go at your own pace - skip what you know.
Foundation - modules 1-5
Install it, find your way around, and learn to talk to it.
Core skills - modules 6-10
Agents, Git, debugging, testing, and a full real-world project.
Going deeper - modules 11-15
MCP servers, skills, hooks and plugins, other languages, APIs, and deployment.
Every module has hands-on challenges at four levels - beginner, intermediate, advanced and expert. Try them yourself first; my solutions are in challenge solutions.
Free resource pack
The course condensed into three PDFs you’ll actually keep open while you work: a 2-page cheat sheet, 5 CLAUDE.md templates, and the 10 prompts I use every day.
The free course teaches you the skills. If you want to go further, I’ve made two paid packs. They’re what I use to build the products above.
Get both as a bundle ($59.99) - saves you $10, and it’s what most people pick. Instant download, updates included. Or buy one on its own:
Advanced Modules ($29.99) - 13 modules, 39+ hours. Production deployment at scale, multi-agent systems, custom MCP servers, enterprise SSO/RBAC, performance, custom agents, sandboxing and plugins, professional workflows, a three-part deep dive into how Claude Code works internally, what’s new in 2026, and using Claude alongside other models.
Real Projects Pack ($39.99) - 14 guided builds you can ship, each with CLAUDE.md templates, rules files, and copy-paste prompts. From an AI todo app and a code-review tool up to a full-stack SaaS boilerplate, an AI agent coordinator, and a multi-agent background pipeline.
评论