Claude Code Getting Started: Setup and First Projects

I started Claude Code the way I start almost every new tool: jump in, no preparation, see what happens. That was a mistake. After a few hours I had a CLAUDE.md so full that Claude was reading dozens of files at the start of every task. Token budget gone, context bloated, Claude began hallucinating. Not maliciously, just because too much was loaded at once and none of it was structured well. The most important tip up front: structure before anything else.

What is Claude Code?

Claude Code is not a chat interface. It runs in the terminal or directly in your IDE and works with your project: reading files, running code, performing Git operations, debugging bugs. All in one session.

The difference from a regular chat assistant is that Claude Code actually works inside your project directory. It sees files, reads them, changes them directly. You don't have to copy code back and forth anymore. That sounds like a small difference, but it isn't: fewer context switches between editor, terminal, and browser save more time than you'd think.

Claude Code currently runs on Claude Sonnet 4.6 as the default model. It's an agentic AI approach: the model plans intermediate steps, executes them, and reacts to the result, instead of just answering your input.

If you want more on the difference between a simple AI assistant and an AI agent: I covered it in more detail here.

Installation: Windows, Mac, Linux

There are two paths: the desktop app (graphical interface, no terminal needed) and the CLI (for use in the terminal or in VS Code). Most people start with the CLI.

CLI: Mac and Linux

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

CLI: Windows

In PowerShell:

irm https://claude.ai/install.ps1 | iex

Prerequisite on Windows: Git for Windows must be installed. WSL is not required.

Desktop app (Windows and Mac)

If you prefer working without a terminal: Anthropic offers a desktop app for Windows and Mac that ships Claude Code with a graphical interface. Linux is currently not supported by the desktop app. The app requires a paid subscription (Pro, Max, Teams, or Enterprise).

Sign in and start

After installation you start Claude Code with claude in the terminal. On first start you're logged in. Two paths: a claude.ai Pro or Max subscription (recommended for regular use), or an API key from Anthropic (pay-per-use, good for testing). With an API key, costs climb fast if your context isn't kept clean. More on that in a moment.

Starting right: structure before the first commit

Before you type claude for the first time, it's worth answering a few questions:

• What do I want to build, and how should the files be organized?
• What should Claude know about the project, and what goes into the CLAUDE.md?

What is the CLAUDE.md?

The CLAUDE.md is a Markdown file in the project directory that Claude reads before doing anything else. It holds the rules for the project: which files are relevant, what naming conventions apply, what Claude should not touch. It's the handover you'd give a new developer joining the project.

The problem: if the CLAUDE.md is too large or references too many files, Claude loads all of it into context for every task. The more tokens in context, the more expensive the session and the higher the chance Claude starts making things up. I've experienced this myself. The fix is not to document more, but to document with more focus.

If context steering interests you, take a look at Context Engineering: challenges and tips. I go deeper there into what happens when models have to process too much at once.

Using Claude Code in VS Code

I use Claude Code exclusively in VS Code, not in a standalone terminal. The reason: I see the file structure right next to the terminal panel. When Claude changes a file, I see immediately which one, and I can check it in the editor in parallel.

The Claude Code extension is free in the VS Code marketplace. After installation, Claude Code opens as a panel inside VS Code and still has full access to the project directory, just like in the terminal.

A tip I use myself: voice input. Instead of typing long prompts, I dictate many instructions. Sounds odd at first, but it works well for complex tasks where I need to give a lot of context. For content marketers or business owners who aren't in the terminal daily, this is a real option.

Three projects from practice

Here are three projects I built with it that actually run.

1. WordPress theme

This website runs on a WordPress theme I built with Claude Code. The 70s theme was my first larger project with it. All of it: templates, block patterns, CSS, PHP functions, web tools. Claude Code didn't just generate code and return it, it worked directly in the files, made changes, and caught errors.

That was also the project where I learned what a good CLAUDE.md looks like and what happens when it gets too big. If you want to see how this looks in practice: I documented the case here.

2. LLM Bot Monitor plugin

The second project was a WordPress plugin: the LLM Bot Monitor. The plugin tracks which AI bots visit my site: crawlers from OpenAI, Anthropic, Google, and others. I wanted to know whether my site is being used as a training data source and which bots come by at all.

Claude Code co-developed the plugin architecture, built the database queries, and wrote the admin interface. I set the direction, Claude handled the implementation, and I reviewed before anything got deployed. Human in the loop was mandatory at every step, especially for a plugin that writes to the database.

3. Carousel automation

I use a skill that automatically generates LinkedIn carousels, a JavaScript-based tool that builds finished slides from input text. I built that with Claude Code: first the logic, then the layout, then the integration as a skill Claude can use itself.

If you want to go deeper on skills, I recommend my Claude Code Skills field report, where I explain what skills are, how they're built, and what works in practice.

Skills and plugins: the ecosystem around Claude Code

Skills are Markdown files that tell Claude how to approach a particular type of task. They're not real programs, just instructions Claude uses as reference while working. A skill file for SEO analysis, for example, can describe which points to check, in what order, and in what format to return the result.

There's a growing community of ready-made plugins and skills. If you don't want to build everything yourself, that's a good starting point.

Skills are too big a topic for this post. There's a dedicated one: Claude Code Skills field report with concrete examples from my work.

Where to start

If you want to learn Claude Code from the ground up, I recommend the course Claude Code in Action from Anthropic. The course covers the basics: installation, first commands, and how to structure projects cleanly from the start. Useful if you don't want to fight through half-finished documentation pages.

Start with a small, real project. Not a proof of concept, but something you actually need. That keeps motivation high and you learn faster what works and what doesn't. Anyone starting with a project that's too big easily gets lost in details Claude Code could solve well if the structure was right from the start.

Voice input is worth trying. For longer prompts where you need to give a lot of context, it's faster than typing.

Prompting well comes with time. The most important point: give context, break tasks into small steps, have Claude check after each step. Not "build me a complete website", but "create the header component to these specs". That sounds like more work, but it's much more reliable. The prompting basics are a good companion read.

Questions, or want to show what you've built? I'm on LinkedIn: eric-hinzpeter-content-marketing.