The way I learned Claude runs in five stages: hold a real conversation first, then build repeatable routines, then let Claude work inside your own files, then build your own roles, and finally let several roles work together. Each stage has a goal you can name, a sign I use to tell that it is behind you, and one mistake that happens often. Starting costs nothing: the free plan covers the whole first stage.
I have used Claude every day since shortly after Claude Code came out, and I have not written a line of code myself in about a year and a half. My business now runs on an AI workforce of roles living in folders on my machine: bookkeeping, content, websites, SEO. What I do with Claude professionally is on my Claude page. This post is about your path there.
What follows is not a course or a curriculum. It is the order in which I learned this myself, and the order I suggest to people who ask me where to start. You are allowed to stop at any stage: most people need stages 1 and 2, and that is a complete result in itself.
The five stages at a glance
Stage 1: Understand Claude and hold the first conversation
What you should be able to do at the end: describe a real task from your working day well enough to get a usable result, and know which plan you actually need.
In practice: create an account, look around the interface, and give it three tasks you would otherwise have done yourself. Not "write me a poem", but the email you have been putting off since yesterday. The step by step is in Claude Guide: Getting Started, what the free plan covers in Claude Free Plan: What You Get. Coming from ChatGPT? The differences are in Claude vs ChatGPT 2026: Compared.
The sign that this stage is behind you: you finished tasks in Claude that were already on your list. Not experiments, things you would otherwise have typed yourself. The bar I suggest: three of them in a week.
The typical mistake here: buying a plan too early. I did not start on the largest plan myself. Buy when a limit annoys you, not before. The plans are compared in Claude Pricing and Plans Explained.
Stage 2: Use Claude in daily work instead of just chatting
What you should be able to do at the end: start a recurring task without explaining your background again, because it is written down somewhere permanent.
If you only use separate new chats with no context written down anywhere, each task largely starts from scratch. The jump to stage 2 is moving context out of your head into something durable: reusable instructions, projects, memory. The groundwork is in Claude Prompts: Templates and How to Use Claude Projects Right. How to check what is actually stored is in Claude Memory: Check It and Use It.
I have used Claude projects myself and never became fully comfortable with them: everything then lives inside Claude, and moving to a different tool is not part of the design. It still works, and it is the easiest route into fixed roles.
The sign that this stage is behind you: you no longer paste background information into the chat. You write the task and nothing else.
The typical mistake here: collecting prompts. A hundred saved phrasings do not replace one page saying how your business works and how you write. The habits that matter here are in Using Claude Well: Six Habits.
Stage 3: From chat to a tool that works in your files
What you should be able to do at the end: give a task inside a folder on your own machine and review the resulting file changes, instead of copying text around.
This is the jump most people postpone for too long. To learn Claude Code is not to learn programming: it is a terminal tool allowed to read, write and run commands, in exactly the folder you point it at. Start with Install Claude Code: Windows, Mac, Linux and Claude Code: A Guide for Beginners. The house rules for that folder are explained in CLAUDE.md: Claude Code's Memory, and the reference is the official Claude Code documentation.
If the terminal puts you off, there is a second route to the same stage: Cowork also works on files, without a command line. I built my own skills and uploaded them there as a zip, which worked well, though I still spend most of my time in the terminal because that is what I am used to. The distinction is in Claude Code or Cowork: What For?.
The sign that this stage is behind you: nothing moves between browser and editor any more. Your last real task changed a file on your disk, and you looked at the change before keeping it.
The typical mistake here: starting without version control. Put the folder under Git before the first larger task. Then changes to your files are visible and usually reversible with one command, and you stop hesitating out of caution. What happens outside that folder, in other systems, Git will not bring back.
Stage 4: Build your own roles
What you should be able to do at the end: describe a recurring task once as a role with a remit, tools and limits, and get repeatable results without explaining the task again.
Three terms separate here that often get mixed up: a skill is an instruction for one task, a subagent works in its own context, and an agent is the role itself. The difference is laid out in Claude Subagents, Skills, Agents, the first role in Build an AI Agent With Claude: 3 Ways.
Two building blocks turn a role into something dependable. Hooks are rules that always apply, whatever Claude happens to be working on: in my setup six events hang off a single shell script that reports every approval request to a local service, so I can decide approvals outside the terminal too. That is described in Claude Code Hooks: Rules That Always Apply. The second are MCP servers, which give a role access to real systems such as your inbox or calendar, set up as in Set Up a Claude MCP Server.
The sign that this stage is behind you: you give the same task on two different days and get the same structure twice, without mentioning the format.
The typical mistake here: writing the role too broadly. "Takes care of marketing" is not a remit. It is a common failure pattern: the wider the description, the more arbitrary the output. I cut roles narrowly for that reason: in my own setup it makes the results more repeatable.
Stage 5: Roles become a workforce
What you should be able to do at the end: let several roles work together, with a handover between them, a separate review, and a point where you decide.
The difference between role and workforce is not headcount, it is organisation. Each of my roles has its own folder with a contract, a personnel file, a file of learned rules and an inbox for handovers. What that means in detail is in What Is an AI Employee?, my own line-up in My AI Workforce: Eight Employees, and the path from an empty folder to a first task in Hiring an AI Employee: The Process.
Two rules matter most. Whoever produces something does not review it: anything that leaves the house is checked by a second role that did not see it being made. And every correction goes into the role's rules file, not into the running chat. Across my workforce that adds up to 312 such rules (as of 14 September 2026), 1,347 lines for Conny, my content editor-in-chief, alone since 20 July.
The sign that this stage is behind you: a correction you gave weeks ago turns up in a new result without you repeating it.
The typical mistake here: having the same role build and review. Checking your own text will not surface your own blind spots. For me that hardened into a house rule: whoever produces something does not review it.
This is where the path stops being general. The questions turn individual, because they now depend on your business rather than on the tool. That is what Claude Practitioners is for, my own German-language community: eight ready-made AI employee packages to load, lessons in the classroom, weekly live calls, and a question and answer area for everything in between. Each package runs an onboarding conversation and asks you itself what it needs to know about your business. Current terms are listed on the community page, and what is inside is described in Claude Community in German: Where?.
How long it takes to learn Claude
I have no sourced figure for this, and I will not invent one. What I can say: stages 1 and 2 run alongside your normal work if you have real tasks every day. Stage 3 costs one focused session to get going and becomes habit after that. Stages 4 and 5 are not a weekend project, they grow with the work. I have been at it for about a year and a half and I am still rebuilding.
What it depends on, in this order:
- Whether you have real tasks. Anyone with something real to finish every day learns faster than anyone practising.
- Whether your work lives in files. Text, spreadsheets and code get you through stage 3 quickly. If almost everything happens in web interfaces it takes longer, because access to the actual data has to be arranged first, through connectors, Claude in Chrome or an MCP server.
- Whether you write your corrections down. That is the lever that did the most for stages 4 and 5 in my own setup.
- How much you have automated before. Anyone who has worked with automation tools already thinks in workflows and handovers.
One reference point for stage 5: loading a finished package, running the onboarding and giving it a first task takes roughly an hour. That is the distance between a loaded package and a first result, not between today and a settled workforce.
Frequently asked questions
Can I learn Claude for free?
Yes, the entire first stage. The free plan is enough to set tasks, judge the answers and find out whether the tool suits you. What it covers and where it stops is in Claude Free Plan: What You Get. You only need a paid plan once the limits slow down your actual work.
Does this work in languages other than English?
Yes. Claude works in German and other languages, and you can write tasks, files and rules entirely in your own language. Mostly English are the official documentation and Anthropic's own learning material. German guides are here on this blog.
Does Anthropic offer learning material itself?
Yes, under the name Claude Academy, free and in English. What exactly is on offer, how it stands on certificates and whether anything exists in German is covered in Claude Courses: What Exists, What's Worth It.
Do I need a course to learn Claude?
Not for stages 1 to 3. The steps are documented, and the real learning is in setting genuine tasks and judging the results. What course offerings exist is in Claude Courses: What Exists, What's Worth It.
Can I skip stage 3?
If you never want to work inside your own files, yes. Your path then ends at stage 2, and that is a complete result. If you want to go further, there is no way around letting Claude write somewhere, whether in the terminal or in Cowork. What Claude Code actually asks of you is answered in Claude Code: A Guide for Beginners.
Do I really have to start in the chat?
I say in talks myself that most people stay in the chat far too long. It is still stage 1 here, because that is where you learn to describe tasks well, and that skill gets used at every later stage. Skip it and you end up at stage 3 in front of a terminal, unsure what to type.
How do I know I am on the wrong stage?
Two signals, measured against this five-stage model. If you are bored and repeating yourself, you are one stage too low. If you are getting results you can no longer judge, you are one stage too high: go back until you can tell good from bad again.
Where to start
Today: create an account and give Claude real tasks from your working day, not test questions. The walkthrough is Claude Guide: Getting Started.
This week: write down the context you currently re-explain every single time. What that looks like is in Using Claude Well: Six Habits.
After that you decide whether stage 2 is enough. If you want to go further, the next step is Claude Code in a folder you have put under Git. And once the questions turn individual, my community has ready-made packages and people walking the same path.
