Claude Code, properly understood

1. Step 1 of 17

   Chat, Cowork or Claude Code?

   Anthropic gives you the same Claude in several places, and the difference is what it is allowed to touch:

   • Claude chat (claude.ai and the Claude apps) talks with you. It works with what you paste or upload into the conversation.
   • Claude Cowork, inside the Claude desktop app, uses the same engine as Claude Code but needs no terminal. It is built for everyday work on your documents and research: it can work with local files, your connected apps and your browser.
   • Claude Code is the one built for working inside a folder or project: reading every file, changing them, running commands and checking its own work. It runs in a terminal, in the desktop app's Code tab, in code editors and on the web.

   Rule of thumb:

   • a question or a draft: chat
   • a task across your documents: Cowork
   • building or changing something made of files (a website, a script, a tidy folder): Claude Code

   Think of it like…

   One expert, three rooms. In chat you talk across a desk. In Cowork they sit at your computer and work on your documents. In Claude Code they are in your workshop with the tools.

   You’re done with this step when

   You can say which of the three you would use to: ask a question, sort a folder of reports, and build a small website.

   I did this

2. Step 2 of 17

   Install the Claude desktop app

   The gentlest way into Claude Code is the Claude desktop app. No typing commands.

   1. Download it from claude.com/download and install it like any other app.
   2. Open it and sign in with your Claude account.
   3. Switch to the Code tab.

   You’re done with this step when

   The desktop app is open, you are signed in, and you can see the Code tab.

   For the curious

   Claude Code also runs in a terminal, in VS Code and JetBrains editors, and on the web at claude.ai/code. The web version works on GitHub repositories in the cloud, not on folders on your computer.

   To install the terminal version on a Mac or Linux:

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

   On Windows, in PowerShell:

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

   Then go into a folder and type claude.

   I did this

3. Step 3 of 17

   Give it a practice folder

   Claude Code works inside the folder you hand it, so start with one where nothing can go wrong.

   1. Make a new, empty folder on your Desktop called claude-practice.
   2. In the Code tab, choose Local, meaning your own computer.
   3. Click Select folder and pick claude-practice.

   Keep important folders out of it until you have finished this guide.

   You’re done with this step when

   A Code session is open on your empty claude-practice folder.

   I did this

4. Step 4 of 17

   Let it make something, and approve it

   Ask for something you can see. Before it writes a file, Claude Code tells you what it wants to do. Read it, then allow it.

   When it finishes, open the folder and double-click the new file. It opens in your web browser.

   See an example: A vague request vs a clear one ▸

   Try this

   Make a one-page website called index.html about my favourite hobby: a title, three short paragraphs and a list of tips. Make it look tidy on a phone too.

   Copy

   You’re done with this step when

   index.html is in the folder and opens in your browser.

   I did this

5. Step 5 of 17

   Change your mind

   You can always say no or go back.

   • If it proposes something you do not want, decline it and say what you would rather it did.
   • If it already made a change you dislike, tell it: "Undo that last change."
   • To stop it mid-task, use the stop button, or press Esc in the terminal.

   Try this

   Change the page to a dark background. Actually, I preferred it before. Undo that last change.

   Copy

   You’re done with this step when

   You saw a change happen and then get undone.

   I did this

1. Step 6 of 17

   The context window: what Claude can see right now

   Claude has no memory of its own between moments. Everything it is working with right now sits in its context window. That means:

   • your conversation
   • every file it has opened
   • the output of every command it ran
   • your standing instructions

   Context is measured in tokens, chunks of text of roughly three-quarters of a word. Current Claude models hold about 1 million tokens, around 750,000 words. That is a lot, but a long session that reads many big files does fill it.

   This is the single most useful idea in this guide. When Claude "forgets" something you said an hour ago, or gets vaguer in a very long session, the context window is why.

   Think of it like…

   A whiteboard. Everything Claude is thinking about must be written on it. When it fills up, something has to be wiped or squeezed down to make room.

   Try this

   /context

   Copy

   You’re done with this step when

   You ran /context and saw what is taking up space in your session.

   For the curious

   What loads automatically at the start of a session:

   • the built-in instructions
   • your CLAUDE.md files
   • a short slice of auto memory
   • the names of connected tools

   Everything else arrives as you work.

   Skills and subagents (chapter 3) exist partly to keep this whiteboard clean.

   I did this

2. Step 7 of 17

   When the whiteboard fills: compact, clear, start fresh

   As the context window nears its limit, Claude Code compacts on its own:

   1. It first clears out old command output.
   2. Then it summarises the earlier conversation.
   3. Your requests and key details are kept.

   Summaries lose detail, though, which is why very long sessions slowly drift. You have two controls:

   • /compact squeezes the history now. You can say what to keep, like /compact keep the colour decisions.
   • /clear wipes the whiteboard and starts a clean conversation in the same folder.

   The habit that saves most frustration: one task per conversation. Finished the website? /clear before starting the spreadsheet.

   Think of it like…

   Tidying a desk. Compacting is filing old papers into a summary folder. Clearing is putting everything away and starting the next job on a clean desk.

   See an example: A tidy working session ▸

   Try this

   /compact keep the decisions about the website design

   Copy

   You’re done with this step when

   You know when you would use /compact and when /clear.

   I did this

3. Step 8 of 17

   Sessions and history: picking up where you left off

   Every conversation is saved as a session, so closing the app does not lose your work.

   • /resume shows your past sessions and lets you jump back into one.
   • /rename gives the current session a name you will recognise later.
   • /rewind (or Esc pressed twice in the terminal) goes back to an earlier point. You can restore the conversation, the files Claude changed, or both.

   Resuming brings back the conversation, so it brings back the full whiteboard too. If an old session got muddled, starting fresh is often better than resuming.

   Think of it like…

   Saved games. You can load an old save, name your saves, and go back to the moment before things went wrong.

   Try this

   /rename hobby-website

   Copy

   You’re done with this step when

   Your session has a name, and you know how to find it again with /resume.

   For the curious

   From a terminal:

   • claude --continue reopens your most recent session in that folder.
   • claude --resume lets you pick one.
   • /branch copies the conversation so you can try a different approach without losing the original.

   Rewind checkpoints cover Claude's own file edits, not everything a command did. Keep real work in Git as well.

   I did this

4. Step 9 of 17

   CLAUDE.md: the house rules it reads every time

   Because each conversation starts with a clean whiteboard, anything you want Claude to always know belongs in a file called CLAUDE.md. It is a plain text note that Claude Code reads at the start of every session in that folder.

   Good things to put in it:

   • what the project is
   • your preferences ("British spelling", "explain like I'm new to this")
   • rules ("never delete files without asking")
   • how to check the work

   Keep it short. Every line takes space on the whiteboard in every session.

   The easiest start is to let Claude write one for you with /init, then edit it.

   Think of it like…

   House rules on the fridge. A house-sitter reads them on arrival, so you do not repeat yourself every visit.

   See an example: A CLAUDE.md for your practice folder ▸See an example: A CLAUDE.md for a small business site ▸

   Try this

   /init

   Copy

   You’re done with this step when

   Your practice folder has a CLAUDE.md, and you added one rule of your own to it.

   For the curious

   CLAUDE.md can live in several places, and they combine:

   • ~/.claude/CLAUDE.md: your rules for every project
   • CLAUDE.md or .claude/CLAUDE.md in a project: shared rules for that project
   • CLAUDE.local.md: your personal notes for one project

   A CLAUDE.md can pull in another file with @path/to/file.md. /memory lists the files in use.

   If a project has an AGENTS.md (the file Codex and some other tools use) and no CLAUDE.md, Claude Code reads AGENTS.md instead.

   I did this

5. Step 10 of 17

   Memory: the notes Claude keeps for itself

   Separately from CLAUDE.md, which you write, Claude Code can keep its own auto memory: short notes about things it learned while working with you, such as how you like things done or a decision you made.

   A small part of those notes is loaded into new sessions, so useful lessons carry over. You can look at and edit them. /memory shows your memory files and opens the auto-memory folder.

   Two rules of thumb:

   • Put rules you insist on in CLAUDE.md.
   • Let auto memory handle preferences it notices.

   Think of it like…

   CLAUDE.md is the rulebook you hand a new colleague. Auto memory is the colleague's own notebook of things they picked up on the job.

   Try this

   /memory

   Copy

   You’re done with this step when

   You opened /memory and know where both kinds of memory live.

   I did this

6. Step 11 of 17

   Permission modes: how much it does without asking

   Claude Code works in one of a few modes:

   • Default asks before each file change or command.
   • Accept edits changes files without asking but still asks about commands.
   • Plan only investigates and proposes a plan, changing nothing. Use it before anything big.
   • Auto lets a built-in safety check approve routine actions and stop risky ones. Recent versions start in this mode in the terminal and in code editors.

   Choose on purpose. While learning, Default or Plan keeps you in the loop.

   Think of it like…

   Trust levels for a new hire: check everything with me, handle routine paperwork yourself, just make a plan, or use your judgement within limits.

   Try this

   Use plan mode: propose how you would add a photo gallery to the page, but do not change anything yet.

   Copy

   You’re done with this step when

   You know which mode you are in and how to switch.

   For the curious

   In the terminal, Shift+Tab cycles through the modes.

   There is also bypass permissions, which skips every check. Never use it outside a throwaway sandbox.

   Rules that pre-approve specific commands live in .claude/settings.json.

   I did this

1. Step 12 of 17

   Skills: teach it a job once

   A skill is a folder with a file called SKILL.md: written instructions for one particular job, such as "how we write release notes" or "how to turn a meeting transcript into minutes".

   The clever part: Claude does not read every skill up front. It sees only each skill's name and one-line description, and opens the full instructions when a task matches. So you can have dozens of skills without filling the whiteboard.

   You can also run one yourself by typing a slash and its name, like /minutes.

   Skills are not only for Claude Code. They work in the Claude apps too on paid plans.

   Think of it like…

   A recipe box. The cook does not memorise every recipe; they see the labels and pull out the right card when someone orders that dish.

   See an example: A skill file: meeting minutes ▸

   Try this

   Create a skill called minutes that turns any meeting notes I give you into: decisions, action items with owners, and open questions. Save it as a personal skill.

   Copy

   You’re done with this step when

   You have a skill, and typing /minutes with some notes uses it.

   For the curious

   Personal skills live in ~/.claude/skills/<name>/SKILL.md. Project skills live in .claude/skills/<name>/SKILL.md.

   The file starts with a short header holding the name and description. The description decides when Claude reaches for the skill, so write it as "use this when…".

   I did this

2. Step 13 of 17

   Subagents: helpers with their own whiteboard

   A subagent is a separate Claude that your main Claude hands a piece of work to. It works with its own context window and sends back only a short report.

   That is the point: if a task means reading fifty files or digging through a long log, the subagent's whiteboard fills up, not yours, and you get the conclusion without the clutter.

   Claude Code already uses built-in helpers like this for searching. You can also define your own specialist, such as a "proof-reader" or a "researcher", with its own instructions and tools.

   Think of it like…

   Sending an assistant to the archive. They come back with a one-page summary instead of carrying every box onto your desk.

   See an example: A subagent file: a proof-reader ▸

   Try this

   /agents

   Copy

   You’re done with this step when

   You opened /agents and can explain why a subagent keeps your main conversation clean.

   For the curious

   A custom subagent is a markdown file at .claude/agents/<name>.md (project) or ~/.claude/agents/<name>.md (personal).

   A header gives its name and description, and optionally which tools and model it may use. The body is its instructions.

   Subagents run as separate requests, so they count toward your usage.

   I did this

3. Step 14 of 17

   Connectors and MCP: plugging into your other tools

   On its own, Claude only sees your folder. Connectors let it reach your other tools: Google Drive, Slack, Notion, GitHub, your calendar and more.

   They are built on MCP (the Model Context Protocol), an open standard for plugging AI into tools. One standard means one plug fits many sockets.

   • In the Claude apps and Cowork, you add connectors from settings.
   • In Claude Code, you can add MCP servers too.

   A connector acts with your access, so only connect what a task needs. Be wary of servers from sources you do not trust.

   Think of it like…

   A universal adapter. Instead of a different plug for every country, one standard socket that any tool can be wired to.

   See an example: What connectors make possible ▸

   You’re done with this step when

   You can explain what a connector does, and you know where to add one.

   For the curious

   In the terminal, claude mcp add adds a server. A project can share its servers in a .mcp.json file. /mcp shows what is connected.

   To keep the whiteboard small, connected tools are described briefly and looked up when needed.

   I did this

4. Step 15 of 17

   Hooks: things that happen automatically

   A hook is a rule that runs on its own at a particular moment. Examples:

   • "after every file edit, tidy the formatting"
   • "when Claude finishes, play a sound"
   • "before any command runs, check it against my list"

   Unlike an instruction in CLAUDE.md, which Claude tries to follow, a hook always runs, because it is your computer doing it, not Claude deciding to.

   Think of it like…

   A doorbell wired to the door. Nobody has to remember to ring it; opening the door does it.

   See an example: A hook: play a sound when Claude finishes ▸

   You’re done with this step when

   You can name one thing you would want to happen automatically every time.

   For the curious

   Hooks are set in .claude/settings.json under hooks. They fire on events such as:

   • PreToolUse and PostToolUse: before and after a tool runs
   • UserPromptSubmit: when you send a message
   • Stop: when Claude finishes
   • SessionStart and SessionEnd
   • Notification

   Each runs a command you choose. Because hooks run commands on your machine, only add ones you understand.

   I did this

5. Step 16 of 17

   Plugins and models

   Plugins bundle all of the above (skills, subagents, connectors and hooks) into one install. That is how a team shares a whole way of working at once. Type /plugin to browse them.

   Models are the different versions of Claude. /model switches between them:

   • Larger models reason more carefully but use your allowance faster.
   • Smaller ones are quicker and lighter.

   Some models can also think before answering. That is slower, but better on hard problems.

   A good default: a capable mid-size model for everyday work, and the larger one when a task is subtle or keeps going wrong.

   Think of it like…

   Plugins are a full toolkit in one case. Models are colleagues with different strengths: one quick, one meticulous.

   Try this

   /model

   Copy

   You’re done with this step when

   You know which model you are using and how to change it.

   I did this

6. Step 17 of 17

   Put it together on a real job

   Pick one small, real task from your week: tidying a folder of downloads, turning notes into a summary, or making a simple page for a club or a small business.

   Use what you learned:

   1. Start a fresh session in a copy of the folder.
   2. Add a CLAUDE.md with your rules.
   3. Ask for a plan first.
   4. Approve the steps.
   5. /clear when you move on to something else.

   Try this

   Look at the files in this folder and suggest a tidier way to organise them. Show me the plan first and wait for my OK before moving anything.

   Copy

   You’re done with this step when

   Claude Code finished a small job you actually needed, the way you set it up.

   I did this

Worth knowing

• It can be confidently wrong. Read what it proposes before you approve it, especially anything that deletes or sends.
• Never paste passwords, bank details or API keys into the chat. If it needs one, put it in a file yourself and tell Claude where it is.
• Everything in a session is sent to Anthropic to be processed, and usage counts toward your plan. Long sessions and subagents use more.

Where this comes from

• Claude Code overview
• Desktop quickstart
• Cowork overview
• How Claude Code works
• The context window
• Sessions
• Memory and CLAUDE.md
• Permission modes
• Skills
• Subagents
• MCP
• Hooks
• Checkpointing

Checked against the sources on 2026-10-08. These tools change quickly.