The Complete Guide to Claude Code (2026 Edition)

Claude Code is Anthropic's terminal agent for software engineering. It edits your files, runs your commands, understands your repo, and ships code with you. This guide is the full reference — from installing the CLI to running a 10-person team on it.

What is Claude Code?

Claude Code is Anthropic's official command-line agent. You type claude in your terminal, and Claude becomes a pair-programmer that can read and write files, run shell commands, search your codebase, and make git commits — all without you copy-pasting snippets between a chat window and your editor.

Unlike Cursor, Copilot, or ChatGPT, Claude Code is agentic. You don't select code and ask for a rewrite — you tell Claude your intent ("add Stripe checkout and wire the webhook") and it plans, edits multiple files, runs tests, and asks you to approve the result. It's closest to a junior engineer who lives in your terminal.

If you're coming from tools like Cursor, the mental model is different. See Claude Code vs Cursor for the full comparison.

How do I install Claude Code?

Claude Code is a Node.js package. You need Node 18 or higher.

# 1. Install globally
npm install -g @anthropic-ai/claude-code

# 2. cd into any project
cd ~/code/my-project

# 3. Start Claude
claude

On first run, Claude Code asks you to sign in with your Anthropic account or paste an API key. Use the account flow — it's easier and token-metered the same way.

Once you're running, try /help to list commands, /resume to continue a previous session, and Esc to interrupt Claude mid-task.

Should I use Claude Code on Windows?

Yes — via WSL2. Native Windows works for simple cases but WSL2 is the supported path for production use. See the free 40-page Claude guide for the full Windows setup walkthrough.

What should my first prompt be?

Do notstart with "build me a login page." Start with a prompt that teaches Claude your codebase. The best opener we've seen:

Read the README, package.json, and the top 3 files in src/.
Then tell me back: (1) what this project does in one sentence,
(2) the main tech stack, (3) what conventions I seem to follow.
Don't write code yet — I want to check you understand the project
before we do anything.

This primes Claude with context and catches misunderstandings in 30 seconds instead of after 40 minutes of bad code. We expand this pattern into 120 tested prompt codes in the Claude Code cheat sheet.

What are Claude Code skills and why do they matter?

A skill is a markdown file that lives in ~/.claude/skills/. When Claude Code starts, it reads every file in that directory and loads them as reusable instructions. If you have react-testing.md in there, Claude knows how you test React components — in every project, forever, without you re-explaining.

Skills are the single biggest productivity unlock in Claude Code. They turn Claude from a generic coder into a senior engineer who knows your exact stack, conventions, and gotchas. Most people don't use them because they don't know they exist — which is why we built a library of 2,300+ free skill files covering React, Next.js, Postgres, Kubernetes, Stripe, OAuth, testing patterns, and more.

How do I install a skill?

# Download a skill from clskillshub.com
curl -o ~/.claude/skills/react-performance.md \
  https://clskillshub.com/skills/frontend/react-performance.md

# Restart claude — the skill is now loaded
claude

Browse skills by category at /browse or grab curated role bundles at /bundles (frontend dev, backend dev, DevOps, full-stack).

How do I write my own skill?

A skill is just a markdown file with a title, a short description, and a set of instructions. The format Claude picks up best is: one topic per file, less than 500 lines, with a "pitfalls" section at the bottom. Use the CLAUDE.md generator as a starting point — it outputs a skill-compatible structure you can split into files.

What are prompt codes (slash prefixes)?

Prompt codes are short prefixes — /skeptic, /l99, /ultrathink, /shipit— that change how Claude thinks before answering. They're not official CLI commands; they're prompt patterns that Claude has been trained (via RLHF and public habit) to respond to.

Examples:

• /skeptic — challenge the premise before answering. Catches wrong questions early.
• /l99— treat this as a senior-engineer question, not a beginner one. Skips "first, install Node.js…" explanations.
• /ultrathink — take extended reasoning time. Best for architecture calls.
• /shipit— remove all placeholders and "[your_name]" style brackets. Finished output, not a template.

We tested 120 of these with before/after examples. The top 11 are free at /prompts and the full set lives in the cheat sheet.

What are MCP servers and when do I need one?

MCP (Model Context Protocol) is Anthropic's open protocol for giving Claude access to external systems — your database, your Notion, your Slack, your GitHub, your browser. An MCP server is a small process Claude Code talks to that exposes tools and data.

You need MCP when Claude needs to do something outside your local files— like read a Notion doc, query Postgres, or open a URL in a headless browser. If your task is "edit this file," you do not need MCP. If your task is "look at our Linear backlog and suggest which tickets to close," you do.

Start with three MCP servers: filesystem (built in), GitHub, and your database of choice. Add more only when a specific task requires it — every MCP server adds tokens to Claude's context, so more is not better.

What are sub-agents and when should I use them?

A sub-agent is a separate Claude instance Claude Code spawns to handle a focused task — like "review this PR" or "write the tests for the function I just wrote." The main agent keeps the overall context; the sub-agent gets a clean window for its narrow job.

Sub-agents are worth using when:

• The task is parallelizable (run 3 test suites at once).
• The task needs a fresh context to avoid bias from the main conversation.
• The task has a clear, verifiable output (lint passes / tests pass / schema valid).

We maintain a set of 10 pre-built Claude Code agents (code reviewer, test writer, security auditor, docs writer, etc.) you can drop into any project. Or build your own with the custom agent builder.

What daily workflows actually work?

After watching a few hundred developers use Claude Code, three workflows consistently produce real output. The rest are demos.

Workflow 1: The "prime then ship" loop

Start every session with a priming prompt (see above), then state your goal, let Claude plan, approve the plan, and only then let it edit files. Planning is cheap; rewriting bad code is not.

Workflow 2: The "skill + code" combo

Before starting a React task, install a React skill file. Before starting a Postgres task, install a Postgres skill file. This sounds trivial but it's the single biggest quality improvement you can make — Claude stops inventing APIs and starts following your patterns. See our curated skill collections.

Workflow 3: The review-before-commit habit

Never let Claude run git commit without you reviewing the diff. Claude is fast and wrong 5% of the time; reviewing a diff takes 30 seconds and catches the bug before it reaches main. The code-reviewer agent can do a first pass before you look at it.

How do teams set up Claude Code?

Running Claude Code on a team of 3+ needs three things in place:

1. A shared CLAUDE.md— the conventions file in your repo root. Every teammate's Claude Code reads this. Use the CLAUDE.md generator to bootstrap it.
2. A team skills bundle — a shared set of team skill filesin a git repo, symlinked into each dev's ~/.claude/skills/ directory.
3. A review rule — no Claude-generated PR ships without a human reviewer. This is the single cultural rule that keeps velocity high and bug rate low.

We've packaged this as a team onboarding pack for frontend, backend, and DevOps teams — same structure, pre-built skills for your stack.

What mistakes should I avoid?

Mistake 1: Not using skills

90% of people using Claude Code have never put a single file in ~/.claude/skills/. They're running Claude in generic mode forever. This is the single biggest waste. Install 3 skills relevant to your stack tonight.

Mistake 2: Asking big questions without /skeptic

Claude will confidently answer the wrong question. Prefix with /skepticwhen the premise might be wrong ("how do I A/B test 200 landing page variants" — Claude will answer, but /skepticwill point out you don't have the traffic for it).

Mistake 3: Letting Claude commit without review

Claude's diffs look right more often than they are right. Always read the diff. This is the #1 source of shipped bugs when using Claude Code.

Mistake 4: Too much MCP

Every MCP server burns context tokens. Start with zero extra servers, add only when a task can't be done without one. Most tasks can.

Mistake 5: Not using sub-agents for reviews

A fresh sub-agent reviewing a PR catches bugs the main agent won't — because the main agent is the one who wrote the code. Use the code-reviewer agent on every non-trivial diff.

FAQ

Is Claude Code free?

The CLI is free to install. You pay for API tokens (or an Anthropic Pro / Max plan). Typical daily use costs $1–$5 for an individual developer.

Does Claude Code work offline?

No — it needs API access to the Claude model. For offline inference you'd use a local model through a different tool.

Can I use Claude Code with GPT-4 or Gemini?

Not directly. If you want a Claude-Code-style experience with a different model, check OpenClaude — open-source CLI with 200+ model support, and every skill at clskillshub.com works with it too.

What's the difference between Claude Code and Claude Projects?

Claude Code is the terminal CLI. Claude Projects is the web UI feature where you attach files to a persistent chat. They're different products. Claude Code is for shipping code; Projects is for long-running research and writing conversations.

Where do I start if I'm brand new?

Three-step path:

1. Grab the free 40-page Claude guide (install, first prompts, skills, prompt codes).
2. Browse the 2,300+ free skills and drop 3 into ~/.claude/skills/.
3. Buy the 120-prompt-codes cheat sheetwhen you're ready to 3x your prompt quality.