Getting started with Claude Code

Claude Code lives in your terminal. No new app to learn, no window to babysit — it sits where your code already is, reads your files, runs your commands, and edits in place. That's the whole appeal: the AI comes to your project instead of you copy-pasting into a chat box.

This is the guide I'd hand someone on day one. Install it, point it at a repo, and learn the five things that make it click. Twenty minutes, start to useful.

Step 1 — Install it

The recommended way is the native installer. It pulls a standalone binary, needs no Node.js, and updates itself in the background.

On Windows it's irm https://claude.ai/install.ps1 | iex in PowerShell. Prefer a package manager? brew install --cask claude-code, or npm install -g @anthropic-ai/claude-code if you want to pin a version. They all install the same binary.

One thing to know before you start: Claude Code needs a paid account — Claude Pro, Max, Team, Enterprise, or an Anthropic Console (API) key. The free Claude.ai plan doesn't include it.

Step 2 — Open a project and say hello

There's no project setup. You cd into any repo and type claude. On first run it opens your browser to log in, then drops you at a prompt.

Start by asking it to explain the codebase, not to change it. It reads files on its own — you don't paste anything. Get a feel for how it explores before you let it edit.

Step 3 — Choose your model

Claude Code ships with three models, and you switch between them mid-session with /model. No restart, no losing context.

The mental model is simple:

• Sonnet (4.6) — your daily driver. It handles the large majority of coding work at a fraction of Opus's cost. Start here.
• Opus (4.8) — bring it out when you hit a wall: architecture decisions, multi-file refactors, the bug that's beaten you twice.
• Haiku (4.5) — fast and cheap for the trivial stuff, where you care about speed more than depth.

There's also opusplan: it reasons through the plan with Opus, then drops to Sonnet to write the code. You get the expensive brain where it matters and the cheap one where it doesn't.

Step 4 — Dial the effort

Models also have an effort level — how hard they think before answering. It's the real speed-versus-quality knob, and most people never touch it.

The levels run low, medium, high, xhigh, max. Low for latency-sensitive busywork, high for non-trivial changes (it's the default on the current models), max for the hardest problem of your week. Set it per session with /effort, or for a single message just drop the word ultrathink anywhere in your prompt. (The old think / think harder phrases aren't special anymore — ultrathink is the one to remember.)

Step 5 — Learn these five basics

Everything above is settings. This is the part that changes how much you get out of it.

1. CLAUDE.md is the memory. Run /init once and it writes a file describing your project. Anything you put there — conventions, commands, "always run the linter" — is loaded into every session. This is the single highest-leverage thing you can do.
2. @ pulls files into context. Type @components/Nav.tsx and that file is on the table. Be specific about what you point it at; context beats cleverness.
3. Plan mode before big changes. Hit Shift+Tab to cycle modes — normal, auto-accept, and plan. Plan mode makes it think the change through and show you the plan before touching a single file.
4. Manage the context window. /clear wipes the slate between unrelated tasks; /compact summarises a long session so it keeps going without forgetting. Cheap habits that keep answers sharp.
5. Set your guardrails. /permissions controls what it can do without asking. Loosen it for a throwaway repo, tighten it for anything that matters.

Where this fits

Claude Code is a tool, and tools are the commodity — the next one will read the same repo just as happily. What carries over is the structure you give it: a clean CLAUDE.md, files worth pointing at, a project that explains itself. Own that layer, and every model that ships next just plugs in.

Install it, point it at something real, and let it explain your own code back to you. That first "oh, it actually gets it" is the moment it stops being a novelty.