Master Claude Code
one stage at a time
An interactive, hands-on guide. No prior coding experience needed — just curiosity and a chat box.
💡 This is a text-based, interactive course. Each stage ends with a quick understanding check and a "try this" prompt you can paste straight into Claude Code.
What is this?
This quest teaches you to work with Claude Code — an AI assistant that doesn't just give advice, it actually builds, edits files, runs commands, and ships projects on your machine.
You don't need to know how to program. You need to know what to ask for. By the end you'll have built a real workspace and be comfortable with memory, permissions, skills, subagents, MCP, and hooks.
What you'll learn
- ✅ Set up a workspace Claude knows inside-out
- ✅ Give it memory you control — and memory it writes itself
- ✅ Build reusable Skills and slash commands
- ✅ Use Git save points, subagents, MCP and hooks
What we won't do
- ❌ Turn you into a software engineer overnight
- ❌ Drown you in flags and config you'll never use
- ❌ Assume anything — every step is explained plainly
💡 Before we start, one honest note
AI tools are powerful but still evolving. Sometimes Claude will need a nudge, a correction, or a second try — and that's completely fine. Learning to guide it, redirect it, and get the best out of it is the whole point. The journey itself is the lesson.
⚠️ Claude Code requires a paid plan. A Claude Pro subscription (~$20/month) or a Claude Max plan is the simplest way in.
Install Claude Code
One line installs everything — or download the app. Whatever you prefer.
What is Claude Code?
Claude Code is a tool where you chat with Claude and it builds things for you. Unlike a normal chatbot, it can create files, run code, and work across whole projects on your computer. You can run it in the desktop app, in a terminal (CLI), or inside VS Code — a free editor that puts your files on one side and a Claude chat on the other. The fastest way to get all of it set up at once is the one-click installer below.
📋 What you need first
- A computer running Windows 10+, macOS, or Linux
- An internet connection
- A paid Claude Pro or Claude Max account
⭐ Option A: One-click installer (fastest)
One line installs everything at once — VS Code, Git, Node.js, and Claude Code — so you skip all the manual steps. It shows a green ✓ for each piece as it finishes.
Windows
Open PowerShell as Administrator (Start → type "PowerShell" → right-click → Run as Administrator), then paste:
irm "https://raw.githubusercontent.com/RanNahmany/claude-code-one-click-install/main/windows/install.bat" -OutFile install.bat; .\install.bat
Mac
Open Terminal (Cmd+Space → type "Terminal"), then paste:
curl -fsSL https://raw.githubusercontent.com/RanNahmany/claude-code-one-click-install/main/mac/install.sh | bash
When it's done, open a new terminal window and type code . to launch VS Code, then sign in to Claude from the sidebar.
A note on trust
This is a community installer by Ran Nahmany. Running a script straight off the internet always deserves a glance first — open the link and skim what it does. If you'd rather install each piece yourself, use Option B or C below.
Option B: The desktop app (no terminal)
Download it
Go to claude.com/download and grab the version for your OS (Windows or Mac).
Install & sign in
Open the downloaded file, install it like any app, then sign in with your Claude account.
Open the Code tab & pick a folder
Switch to the Code tab, click Select folder, and choose a folder for your project. Tip: create a brand-new empty folder so there's nothing sensitive in it.
Option C: The terminal (CLI)
Prefer the command line? Install once with npm, then run claude from inside any project folder:
npm install -g @anthropic-ai/claude-code cd my-project claude
The CLI, the desktop app, and VS Code share the same brain and the same config files. Everything in this guide works in all three.
⚠️ Seeing a "Git is required" error?
If you used the one-click installer, Git is already handled — you won't see this. It only shows up with the manual options (B/C) when Git isn't installed yet. The fix is quick:
Download Git
Windows: git-scm.com/downloads/win. Click the download button.
Install with defaults
Open it and click Next all the way through. You don't need to change a single setting.
Restart Claude
Fully close Claude Code and reopen it. The error will be gone.
💡 On Mac, Git is usually already there. If not, open Terminal and run xcode-select --install.
🤖 Which model should I use?
You'll see a model picker. For learning and most everyday work, Sonnet is plenty — it's fast and smart. Reach for Opus only for gnarly problems where Sonnet gets stuck; it's more capable but burns through your usage faster.
What can Claude Code do that a regular chatbot can't?
Your First Conversation
Claude Code is like a brilliant new teammate. It just needs to know what you want.
🗣️ Just say hello
That's genuinely all it takes to start:
Hi Claude — what can you do in this folder?
🔐 Folder access
Claude will ask permission to access the folder you chose. Before you approve, make sure the folder doesn't contain:
⚠️ Check before you approve
- Password or credential files
- Sensitive personal or business documents
- Banking or payment details
This is exactly why we make a fresh, dedicated folder. Nothing to worry about in a clean folder — approve with confidence.
After you approve, Claude is ready. It doesn't read anything unless you ask — it just waits for instructions.
💡 Tips for a great conversation
- Be specific — "build me a landing page with a contact button" beats "build me a website"
- Give context — tell Claude who you are and what you're trying to achieve
- Ask questions — if Claude does something you don't understand, just ask it to explain
- Have fun — there are no mistakes here, only experiments
What's the best way to talk to Claude Code?
Give Claude a Memory: CLAUDE.md
One file that turns Claude into a permanent expert on your project.
What is CLAUDE.md?
Imagine hiring a new teammate. On day one you hand them a briefing doc: who we are, how we work here, what the standards are. CLAUDE.md is exactly that — a file Claude reads automatically at the start of every conversation in the project.
📄 First — what's a Markdown (.md) file?
Almost every config file here ends in .md. It's just plain text — like Notepad, not Word — with a few simple symbols Claude understands:
That's the whole language. You'll never need more than this.
🧠 Create one in seconds
Just ask Claude to write it for you. Describe the project and how you want it to behave:
Create a CLAUDE.md file for this project. Note that this is a personal blog built with plain HTML and CSS. Rules: keep everything mobile-friendly, use semantic HTML, never add tracking scripts, and always show me a short plan before making more than three file changes.
🔍 What it looks like
From now on, every time you open this project, Claude already knows the rules.
🌍 Two levels of memory
- Project memory —
CLAUDE.mdin the project folder. Applies to this project only. - Global memory —
~/.claude/CLAUDE.mdin your home folder. Applies to every project. Great for personal preferences like "always explain commands before running them."
💡 The 200-line rule
Keep CLAUDE.md short — Anthropic suggests under ~200 lines. Too long and Claude loses focus, like a new hire handed 40 pages on day one. Put the detailed stuff in context/ files (next stage).
What does CLAUDE.md do?
Build Your Workspace
One organized folder that Claude opens — and understands — every single session.
Why a workspace?
Without structure, you re-explain who you are every time. A workspace is one folder, laid out in a way Claude knows to look for. Open it once and Claude already has your rules, your context, and a place to save its work.
🗂️ The structure
🎬 Build it now — one prompt
You don't make these folders by hand. Ask Claude:
Set up my workspace in this folder: create a CLAUDE.md, an empty .env, a .claude/commands/ and .claude/skills/ folder, a context/ folder, and an outputs/ folder. Then tell me what each one is for.
📚 CLAUDE.md = how · context/ = what
This split keeps CLAUDE.md short and your details easy to update. Three files do most of the work:
context/business.md
What the business/project is, who the audience is, the current goal
context/tone.md
Your writing voice — examples of what sounds right and what sounds wrong
context/current.md
What's happening right now — update it weekly
Create context/business.md and interview me to fill it in: ask me what we do, who our audience is, what we never say, and our current goal — one question at a time.
⚠️ One folder, not your whole drive
Open a dedicated my-workspace folder — never your entire C:\ drive. This keeps personal files, passwords, and sensitive documents out of reach.
What's the difference between CLAUDE.md and the context/ folder?
Memory That Writes Itself
You write CLAUDE.md. Claude writes its own notes — and remembers across sessions.
Two kinds of memory
There's what you write and what Claude writes to itself:
- You write → CLAUDE.md: permanent rules and context. You're fully in control.
- Claude writes → MEMORY.md: lessons from your sessions, corrections you gave, decisions you made together. It saves these automatically.
🧷 See it happen
Tell Claude a preference once. It records it and applies it forever — without you repeating it:
Remember that I never want the word "amazing" in anything you write for me — it's not my style. Save that so it sticks across sessions.
Next session it knows. The session after that — still knows. Forever.
🔧 Manage it with /memory
Type /memory and Claude shows what it has saved about itself. You can confirm, edit, or delete any entry. Always in your control. Check it every week or so and clear out what's stale.
⚠️ Auto Memory is local
It lives only on your computer. Anything you want a teammate to also get should go in CLAUDE.md (which you can share), not in Auto Memory.
Who writes MEMORY.md?
Permissions & Safety
Claude asks before it acts. Understanding when — and why — keeps you in control.
Claude asks first
By default, before Claude edits a file, runs a command, or touches anything outside reading, it asks for your approval. You'll see exactly what it wants to do and can Allow once, Allow always, or Deny.
🎚️ Permission modes
You can change how often Claude asks. Cycle modes with Shift+Tab (or set them in settings):
- Default — asks before each new kind of action. Safest while learning.
- Auto-accept edits — stops asking for file edits but still confirms risky commands.
- Plan mode — Claude only researches and proposes; it can't change anything (Stage 10).
🛡️ Set durable rules with /permissions
Type /permissions to manage an allow/deny list. For example, always allow safe read-only commands while always blocking destructive ones:
🔒 Iron rule: secret keys live in .env only
API keys are passwords. Keep them in a .env file — never paste them into chat or hard-code them. If a key ever leaks, go to the service and revoke it immediately, then create a new one.
By default, when does Claude Code make changes to your files?
Slash Commands
Type / to unlock built-in shortcuts — and to create your own.
What's a slash command?
A slash command is a shortcut you trigger by typing / followed by a name. Claude Code ships with many built in, and you can write custom ones for anything you do repeatedly.
🧰 Handy built-ins
/help— list everything available/clear— start a fresh conversation (Stage 11)/init— scan an existing project and generate CLAUDE.md (Stage 9)/review— review your pending code changes/permissions— manage what Claude is allowed to do
✍️ Make your own
A custom command is just a Markdown file in .claude/commands/. The filename becomes the command — post.md becomes /post. Ask Claude to create one:
Create a custom slash command called /standup that summarizes what we changed today as three bullet points I can paste into Slack. Save it in .claude/commands/.
Now typing /standup runs those instructions every time.
Where do custom slash commands live?
Skills: Teach Claude New Abilities
Write down how to do something once. Claude does it every time it's relevant.
⚡ What is a Skill?
Imagine training a new assistant: "When someone asks for a release summary, do this, step by step." You write it down once, and from then on they know. A Skill is exactly that page — an instruction file that teaches Claude how to handle a recurring task.
💡 Slash command vs. Skill — what's the difference?
A slash command runs only when you type it. A Skill can also trigger itself: Claude reads its description, recognizes when it's relevant, and applies it automatically — no command needed.
🔧 What a Skill looks like inside
A simple folder with a SKILL.md file. The frontmatter has a name and a description — the description is what tells Claude when to use it automatically:
🎬 Build one now
Ask Claude to create a Skill for something you do often:
Create a Skill called "changelog" in .claude/skills/changelog/SKILL.md. It should read the recent git commits, group them into Features, Fixes, and Other, and write plain-language release notes to CHANGELOG.md. Give it a clear description so you know when to use it on your own.
⭐ Three Skills worth installing
Superpowers
Forces a disciplined flow — Brainstorm → Spec → Plan → Build → Review — so Claude stops and thinks before rushing in.
Frontend Design (by Anthropic)
Gives Claude a real design sense — deliberate typography, consistent color, polished UI instead of generic pages.
Document Skills (official)
Lets Claude create real Office files: formatted .docx, .xlsx with formulas, .pptx, and .pdf.
⚠️ Before downloading Skills from the internet
A Skill is a text file that gives Claude instructions — a malicious one could tell Claude to do things you didn't expect.
✅ Official Skills
Skills published by the company that makes a product are safe to use.
🤔 Skills from strangers
Just as you wouldn't run software from an unknown source, don't add a Skill from one. The rule of thumb: 20–30 focused Skills beat 100 collected online — every Skill takes up room in Claude's working memory.
How does Claude decide to use a Skill?
Working With Existing Projects
Point Claude at code that already exists and let it get up to speed.
The /init command
When you open a project that already has files and code, run /init. Claude scans the whole project, figures out the structure, and writes what it learned into CLAUDE.md — like giving a new hire an hour to read the docs before they start.
🔎 How to onboard a real codebase
Open the project folder
Select the existing folder in the desktop app, or run claude inside it from the terminal.
Run /init
Just type /init. Claude reads the files and builds a project map in CLAUDE.md.
Ask it to orient you
Now ask questions about the code you inherited.
Give me a tour of this codebase: what does it do, where does execution start, and where would I change the homepage text?
What's the first thing to do with an existing codebase?
Plan Mode: Think Before Acting
For anything non-trivial, let Claude research and propose a plan before it changes a thing.
What is Plan Mode?
In Plan Mode, Claude can read and investigate but not edit. It explores the code, then hands you a step-by-step plan to approve. Nothing changes until you say go — perfect for bigger or riskier tasks.
🗺️ How to use it
Enter Plan Mode
Press Shift+Tab to cycle into Plan Mode, or just ask: "make a plan first, don't change anything yet."
Describe the goal
Tell Claude what you want. It researches and returns a plan.
Review & approve
Read the plan, ask for tweaks, then approve. Claude switches to building.
Plan first, don't make changes yet: I want to add a dark mode toggle to this site. Investigate how styling works here and give me a step-by-step plan with the files you'd touch.
🤔 Push back on the plan
A confident plan isn't always a correct one. Before you approve, ask:
- "Why this approach — is there a simpler way?"
- "What edge cases are you not handling?"
- "What happens if the file is empty or missing?"
If the answer uses jargon you don't follow, say: "Explain that in plain English, no technical terms." Claude will rephrase — always.
In Plan Mode, what can Claude do?
Keep Claude Sharp
Why Claude seems to "get worse" after an hour — and the simple fix.
The context window & "context bloat"
Claude has a short-term memory — a context window of about 200,000 tokens. In a long conversation it fills with noise: old back-and-forth, rejected drafts, corrections on corrections. The result: Claude loses focus and starts to guess. That's context bloat, not a bug.
🧹 The golden rule
A new conversation for every new task. Don't drag a two-hour chat into a totally different topic. Your CLAUDE.md and context files auto-load, so you lose nothing important.
Two commands that keep it sharp
/clear
Fresh start — wipes the conversation but keeps CLAUDE.md and Auto Memory. Use it before each unrelated task.
/compact
Compresses a long conversation — keeps the relevant context, drops the noise. Use it mid-task when answers start slipping.
💡 Analogy: imagine feeding a new worker line after line of changing instructions and "no, not like that." Eventually they lose the thread. /clear hands them a clean page — with your house rules still pinned to the wall.
What does /clear keep when it resets the conversation?
Save Points with Git
Your undo button for everything Claude touches. No Git knowledge required.
Git = version history
Think of Google Docs "version history" — every edit recorded, jump back to any earlier version instantly. Git does the same for your files. Each "commit" is a labeled snapshot. If Claude breaks something, you roll back to the last good one. You're never more than one step from a working version.
💾 Three phrases cover 90% of it
You don't need to learn Git — just say these to Claude:
"Commit this with the message: …"
Saves a checkpoint. E.g. "Commit this with the message: added expense categorization."
"Show me what changed since the last commit"
Lists every modified file, line by line. Review before approving anything big.
"Revert to the last commit"
Instantly undoes everything back to the last save point. Use it when something breaks.
🎬 Try it
Set this folder up with Git if it isn't already, then commit everything as a starting point with the message: "initial workspace".
💡 Good habit
Commit after every step that works — don't wait until the end of a long session. Small, frequent save points = peace of mind.
Claude broke something. What's the fastest way back?
Subagents: Delegate the Work
Claude can spin up helper agents to tackle parts of a job in parallel.
What's a subagent?
A subagent is a separate Claude working on a focused task with its own clean context. The main Claude acts like a team lead: it splits the work, hands pieces to subagents, and combines their results. Great for searching a big codebase, doing several independent tasks at once, or keeping the main conversation tidy.
🤖 Two ways to use them
- Just ask — "search the whole project for where we handle login, in parallel." Claude delegates on its own.
- Custom agents — define specialized agents (a reviewer, a tester, a researcher) as files in
.claude/agents/, each with its own instructions.
Use subagents to do these in parallel: (1) find every place we read environment variables, (2) list all the API routes, (3) summarize the test setup. Then give me one combined report.
💡 Why delegate?
- Speed — independent tasks run at the same time
- Focus — each subagent has a clean slate, so it isn't distracted by the rest of the chat
- Specialization — a dedicated "reviewer" agent reviews better than a generalist mid-conversation
What's the main benefit of subagents?
MCP: Connect Claude to Your Tools
Give Claude safe, structured access to outside services — databases, GitHub, Gmail, and more.
What is MCP?
MCP (Model Context Protocol) is a universal adapter for AI tools. An MCP server exposes a service — like GitHub, a database, Gmail, or Google Drive — and once connected, Claude can use it directly. Think of it as plugging in a new appliance: connect once, use forever.
🔌 Add a server
From the terminal you connect a server with one command. For example, a GitHub server:
claude mcp add github -- npx -y @modelcontextprotocol/server-github
Configured servers are stored in an .mcp.json file. Check what's connected anytime with /mcp. Prefer a no-code route? Tools like Composio connect dozens of apps for you.
🧩 What it unlocks
- "Open a pull request for this change" → via a GitHub MCP server
- "Draft replies to my unread email" → via a Gmail MCP server
- "What were last week's signups?" → via a database MCP server
⚠️ Connect only servers you trust
An MCP server can read and act on real data. Treat it like installing software: stick to official or well-known servers, and never paste secrets where they could be exposed — use environment variables instead.
What does an MCP server give Claude?
Hooks: Automate the Repetitive Stuff
Run your own commands automatically at key moments — without asking Claude every time.
What's a hook?
A hook is a rule: "every time X happens, automatically do Y." It fires on its own — before a tool runs, after a file is edited, or when Claude finishes responding. Unlike Skills (which Claude chooses to use), hooks always run. Like an email rule that auto-files messages from your boss.
🪝 Common events
- PreToolUse — before Claude runs a tool (e.g. block edits to protected files)
- PostToolUse — after an action (e.g. auto-format or run tests after every edit)
- Stop — when Claude finishes (e.g. auto-commit to Git, or verify nothing broke)
⚙️ What it looks like
Hooks live in your settings file. This one runs your formatter after every file edit:
💡 You don't write JSON by hand — describe what you want and ask Claude to set up the hook for you.
Add a hook that automatically commits to Git after every completed task. Write the settings.json config and explain what it does before applying it.
How are hooks different from Skills?
Real-World Wins
Things only Claude Code can do — multiple files, real execution, saved outputs.
Why these need Claude Code
Each task below is impossible in the browser chat — they touch many files at once, run real code on your machine, and save the results. Copy a prompt and adapt the bits in [brackets].
🗃️ Rename & sort 200 invoice PDFs
Read all PDFs in [Invoices]/. Extract vendor, date, and amount from each. Rename to "YYYY-MM-DD Vendor - Description.pdf", create a subfolder per vendor, and move files in. Put anything unreadable in Needs-Review/.
📊 Merge 3 expense CSVs into one report
Read all CSV files in [Expenses]/. Categorize each transaction (software, travel, meals, office, marketing, other). Build an HTML report with subtotals per category, a pie chart, the top 5 expenses, and a grand total. Save to outputs/reports/expenses.html
🌐 Build a landing page and deploy it live
Build a landing page for [business]. What we do: [description]. Audience: [who]. CTA: "Book a free call" → [calendly link]. Style: professional, dark. Deploy it to Vercel and give me the live URL.
Sign up free at vercel.com first; ask Claude to "install the Vercel CLI and help me connect" once, and every future project deploys with one line.
🎙️ Turn 10 call transcripts into insights
Read all transcripts in [Calls]/. Find the top 5 complaints (with direct quotes), top 3 feature requests, and what customers love. Write prioritized recommendations. Save to outputs/customer-insights.md
What makes these tasks need Claude Code instead of the browser chat?
The Finish Line 🏆
Look how far you've come.
You can now…
- ✅ Install Claude Code and hold a productive conversation
- ✅ Give Claude memory you control (CLAUDE.md) and memory it builds itself
- ✅ Set up a real workspace with context and outputs
- ✅ Create slash commands and Skills, and manage permissions safely
- ✅ Keep Claude sharp, use Git save points, and onboard real codebases
- ✅ Scale up with subagents, MCP servers, and hooks
🚀 Where to go next
- Build your workspace for one real project and use it for a week
- Turn your three most repeated requests into Skills
- Add one hook that auto-commits or runs your tests
- Connect one MCP server you'd actually use
📚 Keep learning
Run /help anytime, and explore the official docs at docs.claude.com/claude-code. The best teacher, though, is building something you care about.
💡 Tips & Tricks
Small habits that make a big difference.
🧹 Use /clear often
Starting a new, unrelated task? Clear the conversation. A focused context gives sharper results and avoids Claude mixing up old instructions with new ones.
📅 Tell Claude the date for fresh info
When you want current best practices or recent libraries, mention "as of today" — and ask Claude to search the web if the answer might have changed since its training.
🔁 Let Skills improve themselves
After Claude does a task well, ask it to "save what worked as a Skill." Next time, it starts from your proven approach instead of from scratch.
🧪 You are the QA team
"Tests pass" isn't "it works." Spot-check 3–5 real rows yourself, check a total you can calculate, and be suspicious of output that looks too perfect.
🗣️ Always say "I don't understand"
Any time Claude uses a term you don't follow, stop it: "Explain that in plain language, as if I have no technical background." If you can't follow what it's building, you can't catch its mistakes.
↩️ When it gets messy, start clean
If a task tangles up, don't keep patching. Revert to the last commit and take a fresh approach. A clean redo usually beats a pile of fixes.