Everyone is building amazing things with Claude right now.
And I don’t understand a word of it.
?
CLAUDE.md
Skills
Plugins
MCP
Hooks
Subagents
/Commands
Memory
Checkpoints
Cowork
Routines
Plan mode
/loop
Come on, I’ll explain.
1/36
Let’s start with CLAUDE.md. Why does everyone have one?

CLAUDE.md is basically a user manual that you hand to Claude.

It says how your project works, which rules apply – and what Claude must never do.

CLAUDE.md
Project structure
Rules
Forbidden!
2/36
So I don’t have to explain the same thing every session?

Exactly. Claude reads the file automatically as soon as you start inside the project.

There are even three levels: globally in ~/.claude/, in the project folder and per subfolder. /init writes you a first draft.

CLAUDE.md
Claude
Use this style
Run these tests
Never touch this folder
Claude sticks to the project rules
3/36
Okay. So what are skills, then?

A skill teaches Claude how a particular job is done properly.

Instead of explaining your whole procedure every time, you turn it into a reusable skill.

Procedure
Note
Rule
Template
SKILL
4/36
Give me an example.

Say you want the same analysis report every week.

Your instructions, your procedure, your examples – that becomes a skill you simply call.

Your instructions
Your procedure
Examples
Claude
Skill: weekly report
5/36
And those SKILL.md files everyone downloads from GitHub – they’re just instructions?

Pretty much. A skill is a folder with a SKILL.md in it, plus scripts, templates and reference files.

Claude first reads only the name and description – the rest only when it really needs it. You can also call it yourself with /skill-name.

SKILL.md
scripts/
references/
assets/
A reusable Claude procedure
6/36
And plugins? Aren’t those just skills as well?

Not quite. A skill is one ability.

A plugin is more of a bundle: skills + hooks + subagents + MCP servers + custom commands in one go. You install it from a marketplace with /plugin.

PLUGIN
Skills
Hooks
Subagents
MCP servers
Skill
7/36
Hold on – MCP servers. What are those?

MCP is basically how Claude connects to things outside of Claude.

One standard plug for other people’s programs and services.

Claude
MCP
GitHub
Slack
Database
Drive
Calendar
8/36
So I don’t have to paste everything in by hand any more …

Claude connects on its own, fetches the information – and may act there too, if you allow it.

/mcp shows you what is currently connected.

Without MCP
GitHub
Claude
you copy everything by hand
With MCP
GitHub
MCP
Claude
Claude fetches it itself
9/36
Right. And hooks? That sounds even more confusing.

Hooks are different because Claude doesn’t decide whether they happen.

They fire automatically as soon as a certain event occurs. Set up in settings.json or through /hooks.

Claude
An event happens
automatically triggers the hook
Hook runs
10/36
For example?

A hook can run your formatter after every file change – or stop Claude before a dangerous command runs.

Typical events: PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Stop.

Claude
Claude changes a file
HOOK runs automatically
Format
Test
Check
Report
STOP
Stops before dangerous commands
11/36
What about subagents? Are those literally several Claudes?

They are independent workers Claude can hand clearly defined jobs to – each with its own context window and its own model.

One researches, one reviews code, one tests. Then they report back to the main Claude. You create them as files under .claude/agents/ – or you let Claude do it for you.

MAIN CLAUDE
Research agent
Code agent
Test agent
Results go back to Claude
12/36
And what are all those /something commands everyone uses?

The slash is Claude Code’s shortcut menu.

There are built-in ones like /clear and /compact – and skills bring their own commands that you can call yourself.

/
/clearstart over
/compactboil the context down
/contextwhat eats the space?
/rewindrewind
/modelswitch model
/usagecheck your usage
13/36
Is Claude allowed to do anything at all on my machine?

On the paid plans Claude Code now starts in auto mode: it works straight through, and instead of you a second model reviews every action.

You can switch at any time with Shift+Tab: manual goes back to asking before every change, plan only looks and puts a proposal in front of you.

Autoworks straight through – the reviewer watches, the default
Manualreads on its own, asks before every change
Accept editsmay touch files, everything else still asks
Planonly looks and proposes – changes nothing
switch with Shift + Tab
14/36
Are there shortcuts worth knowing?

Four characters save most of the typing – and two keys are the emergency exit.

Remember Esc above all: you don’t have to wait until Claude is done. Stop it, put it right, carry on.

@filepull a file into the conversation
!commandrun something yourself
/open the command menu
Shift+Tabswitch permission mode
Escstop Claude right now
Esc Escrewind
15/36
Last question: what exactly is this context window?

That is Claude’s working memory for the conversation you are having.

Your input, Claude’s answers, files it has read and tool results all take up room in it.

Context window
Your input
Claude’s answers
Files it has read
Tool results
16/36
So when people say “Claude lost the plot” …

… then usually so much has piled up that it no longer all fits into the working memory at once.

Claude Code then summarises automatically. /context shows you what is taking up the space, /clear gives you a clean start.

CONTEXT WINDOW
Your input
Claude’s answers
Files it has read
Tool results
Your input
Files it has read
Tool results
Your input
What Claude currently has in mind
17/36
And if Claude messes up? Is everything lost?

No – Claude Code sets checkpoints automatically before it changes files.

Esc twice or /rewind, and you roll back: just the conversation, just the files, or both.

State A
State B
State C
this is where it went wrong
← ← ←
/rewind – back to state B
18/36
Does Claude remember anything beyond a single session?

Yes – that is what memory is for: small note files Claude writes itself and reads again the next time it starts.

Your preferences, where a project stands, decisions that keep coming back. Unlike CLAUDE.md, Claude maintains this itself.

Claude
MEMORY.md
note.md
survives from session to session
19/36
And all of this only runs in the terminal?

Not for a long time now. Claude Code comes as a terminal command, as a desktop app, in the browser at claude.ai/code and as an extension for VS Code and JetBrains.

On top of that, agents that keep working in the background or in the cloud while you do something else.

Terminal
Desktop app
Browser
VS Code / JetBrains
Claude
the same session, everywhere
20/36
Is there a version of all this to click on?

Yes. The Claude desktop app has three tabs: Chat is the conversation, Cowork does office work, Code is Claude Code with a user interface.

It is available for Mac and Windows, and as a beta for Linux.

Chat
Cowork
Code
Talkingquestions, drafts, ideas
Delegatingdocuments and folders
Developingsource code and Git
21/36
What can the app do that the terminal can’t?

Several sessions side by side – each in its own Git worktree. That way they don’t get in each other’s way, even when they work on the same spot.

Plus editor, terminal, preview and the review of the changes in the same window. You can even kick off a session from your phone.

Session 1its own worktree
Session 2its own worktree
Session 3its own worktree
Editor
Terminal
Preview
Review changes
everything in one window
22/36
And what if a change runs through the whole project?

That is what /batch is for: Claude thinks the rebuild through once, then hands it out to many agents that start at the same time – each in its own working copy.

At the end each one puts up its own set of changes. You see them side by side and accept them one at a time.

think it through once, then split it up
Agent 1its own copy
Agent 2its own copy
Agent 3its own copy
and so onup to thirty
one set of changes each, to review
23/36
So what is Cowork, then?

The same machinery as Claude Code – only in the ordinary Claude app instead of the terminal, with nothing to set up.

You share folders, describe the goal, and Claude gets going: spreadsheets with real formulas, presentations, finished documents.

Your folders
Your accounts
COWORK
Spreadsheet
Presentation
Report
24/36
So what does it actually do?

Turn a folder full of receipt photos into an expense report. Write a report from a pile of notes. Tidy up a filing system. If need be it will drive Chrome for that – clicking, typing, filling in forms.

How much it may decide on its own is up to you.

Receipts → expense report
Notes → report
Tidy up the filing
and in the browser if that’s what it takes
Asks every timeyou approve everything
Decides for itselfwith a safety net
Just gets on with itwithout asking
25/36
When do I use which?

The question is always how far Claude may go into your things. Cowork gets individual folders, runs sealed off and needs no setting up.

Claude Code gets the whole machine. Only here can you teach it tools of your own – and only here does it keep going when you are away.

1 · In the conversationnothing to set up – you upload, you download
2 · Coworkdesktop app: Claude works in your folders
3 · Claude Codeyou are here: the whole machine, your own tools, runs without you
getting things done → Cowork · building tools that run by themselves → Code
26/36
And what if something should run regularly on its own?

Then you create a routine: task, project and accounts saved once – after that it runs without you.

And it does so in the cloud. Your machine may be switched off. You create it with /schedule, in the app or in the browser.

/schedule review the new pull requests daily at 9am
Task
Project
Accounts
runs in the cloud – machine off
27/36
And what sets a routine off?

Three things, alone or combined: a schedule, a call from outside, or an event on GitHub.

The shortest interval is one hour. One-off appointments work too – /schedule in two weeks … The whole thing is still in preview.

Schedulehourly to weekly
A callfrom your own tools
GitHubnew pull request
Routine runs
the result is waiting in the morning
28/36
And if it’s only meant for the next hour?

Then use /loop. It repeats a task inside the open session – on your machine, with your files.

Leave out the interval and Claude picks it itself: short while something is happening, longer when things are quiet. Esc ends the loop.

/loop 5m check whether the build has finished
↻ ↻ ↻
with an intervalevery 5 minutes
without oneClaude decides
with no task at allit tidies up by itself
runs only while the session is open – and ends after seven days
29/36
That sounds almost like a routine.

The difference is where it runs. The loop needs your session open and your machine switched on – in return it sees your files.

The routine runs in the cloud, without you. Rule of thumb: loop for this afternoon, routine for every Monday.

/loop
on your machine
session has to stay open
sees your files
for right now
/schedule
in the cloud
machine may be off
a fresh copy of the project
for every week
30/36
And what if I don’t know yet what it should look like?

Then start with /design. Claude lays several drafts out as boards on a canvas – click them, move them, edit the text right there.

Once the direction is right, you hand it back to Claude Code, and that is where it gets built. Still an early preview.

describe it
look at drafts
adjust it yourself
have it built
several proposals side by side
click instead of describe
still an early preview – may change
31/36
And which model is actually doing the work?

As of August 2026: Fable 5 is the strongest and built for long runs, Opus 5 takes the heavy lifting, Sonnet 5 the everyday work, Haiku 4.5 the fast and cheap jobs.

Switch with /model. With /fast Opus answers more quickly – it is not swapped for a smaller model in the process.

Fable 5long agent runs1M context
Opus 5the heavy jobs1M context
Sonnet 5everyday work1M context
Haiku 4.5fast & cheap200K context
switch with /model
32/36
And how thoroughly it works – can I say that too?

Yes, that is the second dial next to the model: /effort. First pick the model by how hard the task is, then the effort by how thorough you want it.

The default is high. Going down means faster and cheaper, going up means more thinking, more tool calls, longer runs.

lowshort, simple jobs – the fastest
mediumthe thrifty middle ground
highthe default – for anything demanding
xhighagent runs lasting hours
maxeverything it has, never mind the spend
set it with /effort
33/36
And how do I best put it?

State the goal, not the individual steps – Claude finds the way itself. And say how you can tell when it is done.

And give the context: who it is for, what for, and what must not happen under any circumstances. The clearer the task, the fewer rounds.

rather not
“Make the table look nicer.”
better
“The table goes to the board. Numbers right-aligned, totals row in bold, no colours. Check at the end that the totals add up.”
34/36
Can I actually rely on it?

Usually yes – but Claude can also be convincingly wrong. Have it show you what it did instead of just believing it.

And whatever can be checked, let Claude check itself: run the tests, redo the arithmetic, name the sources.

Look at the changesnot just the summary
Run the testsautomatically if you can
Ask for sourceswhere does that number come from?
When in doubt, rewindEsc Esc costs nothing
Trust is fine. Checking is faster than repairing.
35/36
And what if a file says “delete everything”?

Good question – that is exactly where the trap is. Claude reads web pages, e-mails, tickets and other people’s source code. None of that is an instruction from you.

Instructions come from you, from your project rules and from what you allow. Everything else is material to read – which is why the permissions are not there to annoy you.

Your instructioncounts
CLAUDE.mdcounts
Web pagematerial only
E-mailmaterial only
someone’s ticketmaterial only
The further out Claude may reach, the shorter you keep the leash
36/36
Ohh. I think I’ve got it now.

Glad to hear it.

💡
CLAUDE.md – the rules of your project
Skills – how a job is done properly
Plugins – all of that as one bundle
MCP – the connection to the outside
Hooks – run automatically on events
Subagents – helpers with a mind of their own
Context window – the working memory
Checkpoints & memory – rewind and remember
Cowork – the same inside your folders, without a terminal
Routines – run to a schedule in the cloud
Permissions – automatic today, stricter whenever you want
Effort – how thoroughly the work should be done
/loop & /batch – repeat it, or split it across many
/design – draft it first, then have it built