I've been building with AI tools for a while now. ChatGPT, Cursor, Copilot, Gemini — I've put real hours into all of them.

But the one that changed how I actually work — not just how I chat, not just how I autocomplete code — is Claude Code.

It's the tool behind the AI growth team we built at Superpower (Part 2) and the personal AI chief of staff I run 24/7 on a $5 server (Part 3). But before I get into those builds, I want to talk about the tool itself. Because most people I show it to have the same reaction: "Wait, it can do that?"

Claude Code is not a chatbot. It's not an autocomplete engine. It's an AI that lives in your terminal (or your IDE) and can do things.

It reads your files. It understands your project structure. It writes code, edits code, runs commands, searches the web, talks to external APIs, creates files, and takes actions — all from a single conversation.

Think of it like this: ChatGPT is a brain in a jar. You paste things in, it thinks, it responds. Claude Code is a brain with hands. It can see your project, reach into your files, and make changes directly.

The first time I used it, I described a bug I was seeing. Instead of giving me a code snippet to copy-paste, it read the relevant files, found the issue, and fixed it. In place. In my actual codebase. I just had to approve the change.

That's a fundamentally different interaction model. Anthropic's engineering team wrote up how they think about it — worth reading if you want the mental model from the people who built it.

Here's the most important concept in Claude Code, and it's embarrassingly simple.

You create a file called CLAUDE.md at the root of your project. In it, you write instructions. Plain English. Markdown. That's it.

Whatever you put in that file, Claude reads it at the start of every conversation. It becomes the AI's operating manual for your project.

Ours at Superpower includes:

But you don't need to start there. Your first CLAUDE.md could be three lines:

You can even get claude to update itself by just telling it to. I have never actually editted a single line in any of the files myself. If you want to go deeper on this, the official memory docs explain how Claude reads CLAUDE.md files recursively from your project root. And this guide from HumanLayer has practical tips on structuring them well (their key advice: keep it under 300 lines and make every line universally applicable).

The second concept worth understanding is MCP — Model Context Protocol.

In simple terms: MCP lets Claude Code connect to external services. Notion, Figma, Slack, Klaviyo, PostHog, databases — anything with an API.

You create a .mcp.json file and point it at the services you want. Once connected, Claude can read and write to those services through natural language. You say "create a page in Notion with this content" and it does. The official MCP docs explain the protocol in detail, and there's a registry of public MCP servers you can browse — think of it like an app store for AI integrations.

Anthropic also has a free MCP course if you want to build your own servers from scratch. But honestly, for most people, the pre-built ones cover 90% of use cases.

I'm not going to get deep into the technical setup here — that's what Parts 2 and 3 are for. The point is: Claude Code isn't limited to your local files. It can reach out into your whole tool stack.

Here's how I think about it.

Most AI tools are reactive. You ask, they answer. You paste, they respond. The AI is a consultant you keep calling into meetings.

Claude Code is operational. It sits inside your workspace. It knows your files, your conventions, your tools. You don't paste context — it already has context. You don't describe your project — it can read your project.

The shift is from "let me explain my situation to the AI" to "the AI already knows my situation — let me just tell it what I need."

That might sound like a small difference. It's not. It's the difference between having a smart friend you call for advice and having a capable teammate who sits in your codebase every day.

A few real examples from the last week:

"Read my last 5 newsletters and update the style guide" — Claude pulled articles from Notion, analyzed patterns in voice, structure, and vocabulary, then updated three style guide files. Took about 2 minutes.

"Maya, create Meta ad concepts for women with fatigue" — Maya is one of our AI agents (more on this in Part 2). Claude became Maya, ran Reddit research, generated 30 ad concepts, wrote briefs, produced copy, and flagged compliance issues. One prompt.

"Build a sponsorship pipeline in Notion with these 12 companies" — Claude created the database, set up the properties, loaded the data, and built the tracking page. I described what I wanted in plain English.

None of these required me to write code. None required me to learn a special syntax. I described what I wanted, and the AI figured out how to do it.

That's the pattern. Set it up once. Then just talk to it.

If you've never used Claude Code, here's the fastest path:

That's it. You don't need to understand plugins, agents, MCP, or any of the advanced features to start getting value. Those come later, when you're ready to build the kinds of systems I'll cover in the next two parts.

If you want a more structured intro, Anthropic has a free course called Claude Code in Action that walks through real projects step by step. It's good. But honestly, the fastest way to learn is to open it in a project and start asking it to do things.

If you want to go deeper, here's what I'd actually recommend:

Official

Community & Guides

You don't need all of this to start. Install it, create a CLAUDE.md, and go. Come back to these when you're ready to level up.

This is a three-part series:

Part 2: How to Build an AI Growth Team on Claude Code

How we built 15 AI agents, 19 modular skills, and a shared knowledge library at Superpower — all inside a single Git repo. Agents with names, roles, and workflows that the entire team uses daily.

Part 3: How I Built My AI Chief of Staff

How I built a personal AI operator that monitors Slack, reads my calendar, tracks revenue, and messages me on Telegram. Runs 24/7 on a $5/mo server. Plus: how I used ChatGPT to write the personality file (and why it was better than anything I'd have written myself).

Both builds start from the foundation covered in this article. CLAUDE.md is the backbone. MCP is the nervous system. The rest is just markdown files in folders.

And that's enough.

A few tools and collections I’ve built (or obsessively curated) over the years:

These are tools and services I use personally and professionally — and recommend without hesitation: