How to learn Claude Code: four steps, in this order

By Marco Kohns, Co-founder of ENLIX, lecturer in AI and growth

· 11 min read

You learn Claude Code in four steps: install it and give it a task that matters, cut tasks so they run end to end, write your own rules file, and only then Skills, MCP servers and agents. That order is not the order the features are documented in, which is why so many first attempts get stuck in step four before step three has started.

Documentation sorts by feature surface. It has to, because it describes a product. Learning works differently: you need whatever is in your way on a Tuesday afternoon first. So what follows is not a feature list but an order, with a stop criterion for each step, meaning the sentence that tells you it is time to move on.

In what order should you learn Claude Code?

In the order a working project needs things, which differs from how the docs are laid out. Anthropic's own self-paced course Claude Code 101 shows this itself: twelve lessons take you from "What is Claude Code" through installation and the daily workflow to a final block called "Customizing Claude Code", and only there do CLAUDE.md, subagents, Skills, MCP and hooks appear, all five side by side.

Five topics in a closing block is correct for a syllabus. As a learning order it misleads, because one of those five comes up in daily work long before the other four, and three of them you do not need in your first weeks at all.

The same topics, two orders

step that leaves a file behind

How the features are sorted

  1. Installation
  2. First prompt
  3. Daily workflow
  4. Context management
  5. CLAUDE.md
  6. Subagents
  7. Skills
  8. MCP
  9. Hooks

1 of 9 steps leave a file behind

How a project needs them

  1. Install and give it a task from your own work
  2. Cut tasks until they run end to end
  3. Write a CLAUDE.md as soon as you repeat yourself
  4. Skills and MCP once a task recurs

3 of 4 steps leave a file behind

Left lane: the lesson order of Claude Code 101, retrieved 19 October 2026. Right lane: the order in which this project needed the topics.

The difference between the lanes is not the content, it is what you are holding at the end of a step. On the left you read eight topics and write one file. On the right, from step two onward, every step produces something that stays in the project.

Step 1: What has to be in place before the first prompt?

A terminal, a project directory and an account. Nothing else, and in particular no prepared collection of context. Anthropic says so explicitly in the quickstart: "Claude Code reads your project files as needed. You do not need to add context manually" (Quickstart).

Installing is one command. On macOS, Linux and WSL:

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

On Windows PowerShell it is irm https://claude.ai/install.ps1 | iex, on CMD the install.cmd variant. After that, claude --version confirms a version number followed by (Claude Code). The places people get stuck anyway, above all the PATH on Windows, are worked through one by one in installing Claude Code. Which of the four entry points suits you is decided in Claude Code CLI; at the start it is whichever one you already have open.

The first prompt belongs in a project that is yours, not in a practice repository. The docs suggest questions that change nothing, such as "what does this project do?" or "explain the folder structure". This is the one case where a question is worth more than a task: you get to see what the program reads out of your files before it touches them.

Stop criterion: you have accepted a change the program proposed, and you know which key switches permission modes. In a running session that is Shift+Tab.

Step 2: How do you cut a task so it runs end to end?

By describing it so the result can be checked. That is the whole step, and it is both the shortest and the most underrated one. The docs show the difference with a pair of examples: not "fix the bug", but "fix the login error where users see a blank screen after entering invalid credentials".

The same rule in three shapes worth memorising:

Instead ofBetterWhy
"Make this nicer""Set the heading to 2 lines of spacing"Checkable without asking you
"Write tests""Write a test for the empty cart"One result instead of a direction
"Clean this up""Delete the unused imports in src/lib/"One place, one operation

Multi-step work gets numbered rather than packed into one sentence. The quickstart writes three lines underneath each other rather than a paragraph, and that is also how it processes better. The four-block structure we teach for this is in prompt engineering, and the prompt generator assembles it from your task for free.

Stop criterion: one of your tasks has run from start to finish once, without you correcting anything midway.

Step 3: When do you write your first CLAUDE.md?

The moment you repeat yourself for the second time. This is the step where learning stops being operation and starts being work on a file. Every session starts from zero: "Each Claude Code session starts with a fresh context window" (project memory). The CLAUDE.md file is what comes along anyway.

/init takes the first pass off your hands. It analyses the codebase and generates a starter file with build commands, test instructions and the conventions it finds. What it cannot do is the actual point: writing down the instructions it would never discover by itself.

How much that turns into is something you can check on this website. It runs out of a repository whose CLAUDE.md we counted for this article:

This project's rules file, lines per section

Architecture260Design65Hard rules33What this is not14Open items10Gotchas8

Counted on 19 October 2026 in the CLAUDE.md of academy.enlix.ai: 431 lines, 3,836 words, 7 numbered hard rules. Anthropic states a target of under 200 lines per file.

This project's rules file, lines per section
Architecture260
Design65
Hard rules33
What this is not14
Open items10
Gotchas8

Two things about that are instructive, and the second one is uncomfortable.

First: the largest section is architecture, not prompt technique. Two thirds of the file describe how this project is built and which decision was made why. So what you learn while writing a CLAUDE.md is not how to phrase things better, it is how to describe your own project more precisely.

Second: 431 lines is more than double what Anthropic recommends. The documentation states a target of "under 200 lines per CLAUDE.md file" and gives the reason: "Longer files consume more context and reduce adherence." The documented way out is path-specific rules under .claude/rules/, which load only when Claude works on matching files. Imports via @path help you organise, but not with the context cost, because imported files load at startup too. That split is still outstanding on our side, and it is the next piece of work on that file.

Every one of those 7 hard rules was written after the fact, once something had gone wrong. That is not a planning failure, it is the shape this step takes: you write down the rule that cost you time last time. Since version 2.1.283 you can also have the file reviewed. /doctor prompt-audit looks for outdated and contradictory instructions and changes nothing until you ask it to.

Stop criterion: an instruction you used to type every time now sits in the file and gets followed without you mentioning it.

Step 4: When are Skills, MCP and agents worth it?

Once a task recurs, not before. The three topics are the extension layer, and they divide cleanly by what they solve:

  • Skills are written-down workflows that load when the occasion fits. They answer "how do we do this here". What that looks like, and what belongs in the CLAUDE.md instead, is in Claude Code Skills.
  • MCP servers connect the program to systems outside your directory so it reads and writes there instead of making you copy. Which two we run is in Claude Code MCP.
  • Agents carry out multi-step work without you, on a schedule. What holds up and what we switched off again is in what is an AI agent.

The order inside this step comes from your work, not from a ranking. Someone building the same report every day needs an MCP server. Someone writing the same kind of text every day needs a Skill.

Stop criterion: there is none here. This step does not end, and that is the difference between "I can operate it" and "I run things with it".

What can you skip at the start?

More than most introductions admit. The guides on this topic are thorough; one of the most detailed German ones runs to 18 sections and a free online course to 14 chapters. Completeness is right for a reference work and gets in the way during your first week.

These four topics can wait in good conscience until you need them:

TopicPostpone untilReason
Hooksa mistake must not happen againHooks enforce, a CLAUDE.md only asks
Subagentsa task is too big for one sessionBefore that it is admin without benefit
Token-saving tricksthe bill starts to bother youThe cost calculator answers that in a minute
Model selectionyou can tell a differenceThe default is right for getting started

The distinction on hooks is the most important one, and it is rarely put as plainly as the documentation puts it: Claude treats CLAUDE.md files "as context, not as enforced configuration". Anyone who truly wants to block an action needs a PreToolUse hook, not a more sharply worded rule. That, however, is an answer to a problem you do not have yet.

How long does this take?

The first two steps are a matter of hours, the third a matter of weeks, and the fourth does not end. Anthropic's own course budgets 1.5 hours for the basics, spread over twelve lessons. That matches what steps 1 and 2 above ask for.

For the third step there is no hour count, because it hangs on your project rather than on lessons. We moved it forward in our own curriculum for that reason: the operating-system module sat in week 4 on 14 September 2026 and moved to week 1, because it is the on-ramp and not the send-off. In the same revision, data protection came out of the first version, because no module stood behind it and a promise with nothing in it is worse than a gap.

Modules per week in the course curriculum

Week 13Week 22Week 32Week 42

Four weeks, four live sessions of 45 minutes. The four week blocks on the course page subdivide into nine modules in the course area, 3/2/2/2. Source: our own course registry, as of 19 October 2026.

Modules per week in the course curriculum
Week 13
Week 22
Week 32
Week 42

That the first week is the widest is no accident and follows directly from the same reasoning: the on-ramp carries three modules, the three weeks after it two each.

Do you need a course for this?

Not for steps 1 and 2. The documentation is public, Claude Code 101 is freely accessible, and the path to your first task running end to end is short enough that a guide is sufficient. Anyone who backs themselves to do that should start there and nowhere else.

What a course shortens is steps 3 and 4. Otherwise you derive a rules file and an extension layer from your own mistakes, and that takes exactly as long as it takes. That is what The Claude Code System contains: four weeks, nine modules, with a finished operating-system template in week 1 that you rewrite for your own business instead of working it out. The cohort runs in German.

To be clear about it: if all you want to know is whether the tool is for you, no payment answers that. The afternoon from steps 1 and 2 does. And if the question before the payment is whether funding applies, two approvals decide it, not the course description. And if the choice is between several providers, the checked prices and formats are in the Claude Code course comparison. And if you are starting that afternoon today, it is written out move by move in the guide to the first session.

Frequently asked

In what order should you learn Claude Code?

Install it and give it a task from your own work, learn to cut tasks so they run end to end, write your own CLAUDE.md, and only then Skills, MCP servers and agents. The documentation is ordered differently: it sorts by feature surface, not by what you need first.

How long does it take to learn Claude Code?

The first two steps take an afternoon. Anthropic's own course Claude Code 101 states 12 lessons and 1.5 hours. The third step, the rules file, does not end. Ours has grown to 431 lines, and every rule in it is there because something went wrong first.

Do you need to know how to code to learn Claude Code?

Not to get started. You need a terminal and a project directory. What you learn instead is how to describe a task and how to check a result. Anthropic puts it this way in the quickstart: talk to Claude like a helpful colleague.

What is a CLAUDE.md file and when do you need one?

A Markdown file in your project directory holding instructions that load at the start of every session. You need it the moment you repeat yourself for the second time. `/init` generates a first version from your codebase; refining it is your job.

Do you need a course to learn Claude Code?

Not for steps 1 and 2. The documentation is public and Anthropic's Claude Code 101 is a self-paced course. A course shortens steps 3 and 4, because it hands you a finished structure instead of making you derive one from your own mistakes.

Written by

Marco Kohns

Co-founder of ENLIX, lecturer in AI and growth

Marco worked as a growth product manager at a Silicon Valley scale-up and has been teaching that way of working ever since. Today he runs ENLIX with Tobias and builds two products of his own on the same systems, which is what the courses open up.

  • Growth product manager at a Silicon Valley scale-up, Series A to B, backed by a16z, General Catalyst and Sapphire, with users in over 100 countries and more than 20,000 cities
  • Peer-reviewed research in the Journal of Business Research on generative AI in growth, with Prof. René Bohnsack. The research began in summer 2022, months before ChatGPT was public
  • Executive education lecturer at Católica-Lisbon, over 10 seminars, more than 1,500 people taught

Keep reading