You need the right mental model before the first command. Get this wrong and nothing makes sense. Get it right and everything clicks.
What People Think It Is
Most people approach Claude Code as a very good chatbot that can also write code.
They type a question. They get an answer with a code block. They copy the code somewhere. They try to run it. Something breaks. They paste the error back. They get another block. They copy it again.
This workflow is exhausting, fragile, and misses everything that makes Claude Code actually useful.
What It Actually Is
Claude Code is an agentic system operating inside your terminal.
Here's what that means in practice:
When you give it a task, it doesn't produce a single response and wait. It takes action. It reads your files. It writes code directly to your project. It runs commands. It reads the output. It fixes what broke. It makes decisions -- within the scope you've defined -- and reports back when it needs direction or has finished.
You are not copy-pasting. You are directing.
The analogy that works: Claude Code is closer to a skilled contractor than a consultant.
A consultant talks. A consultant hands you documents. You implement what the consultant recommends, alone, with the documents.
A contractor shows up, assesses the job, does the work, and asks when they need a decision from you. You're present. You're directing. But you're not doing the work yourself.
Claude Code is the contractor. The work is building software.
The Three Modes of Operation
Understanding these will save you significant frustration:
1. Conversation mode
You and Claude Code discuss what to build. You describe requirements. It asks clarifying questions. This is planning, not building.
2. Build mode
You give a specific task. It executes -- reads files, writes code, runs tests, handles errors. You watch and intervene when needed.
3. Review mode
The work is done. You review what was built. You ask it to explain decisions, adjust behavior, or fix specific issues.
Most productive sessions move through all three. Most unproductive sessions stay stuck in conversation mode, never actually building anything.
What "Agentic" Means for You
The word "agentic" means the system takes initiative within defined boundaries.
This is different from a tool that waits for your exact instructions at every step. Claude Code will, within the scope of your task, make reasonable decisions autonomously and tell you about them.
Example: you ask it to build a CSV export for your dashboard. It will:
- Choose an appropriate file format and structure
- Handle edge cases (what if a field is empty?)
- Name the file sensibly
- Place it in a logical location in your project
It won't ask you about each of these decisions. It will make them and tell you what it chose. You can override any decision, but you don't have to weigh in on all of them.
This is an efficiency. It's also a responsibility: you need to review what it built, not just whether it ran.
What It Cannot Do
It cannot read your mind. Vague descriptions produce vague results. "Make it better" is not a task. "The report takes 4 seconds to load and it should be under 1 second" is a task.
It cannot guarantee correctness. Code that runs is not code that's right. You need to test that the output is actually what you intended. Claude Code will help you write tests, but it won't verify that your business logic is sound -- only you can do that.
It cannot fix problems it doesn't know about. If there's context about your business, your data, your constraints -- say it explicitly. Don't assume it knows.
It cannot replace judgment. It will make architectural choices. Some will be excellent. Some will be wrong for your specific situation. You need to stay present enough to recognize the difference.
The Key Mindset Shift
You are not asking Claude Code to do something for you.
You are building something with it.
The distinction matters because it determines how you engage. If you treat it as a service -- give instruction, receive output, move on -- you will build things you don't understand. When they break (they will), you'll be helpless.
If you treat it as a collaborator -- stay present, ask questions, understand decisions, push back when something seems wrong -- you'll build things that last and that you can maintain.
Tobi Lütke didn't become successful because he found a tool that built Shopify for him. He became successful because he understood every layer of what he built. The tool changed (from hand-written Ruby to AI-assisted development). The standard didn't.
Keep the standard.
Next: Setup in 20 Minutes