Claude Code for beginners: where to start if you're not a programmer

Claude Code is Anthropic’s AI agent for working with files and programs on your computer. You describe the task in plain words, and it reads documents, makes edits, runs tools and checks the result. It works in the terminal, in the Code tab of the Claude desktop app and in the browser. The simplest way to start is the Pro plan at $20 a month. You don’t need to know how to program to get going.

Below is a short route from “installed it, opened it, now what?” to the first real task you hand over to the agent in full: nine steps, checked against Anthropic’s documentation, with commands, prices and risks.

What Claude Code is and how it differs from a chatbot

The Claude chat in the browser can also create files and run code, but it works in its own sandbox: you upload the documents, and you collect the result yourself. Claude Code in the terminal, in a code editor or in the desktop app in Local mode works directly with your folders and programs. It takes the task, breaks it into steps itself, opens the files it needs, writes, checks and brings back a finished result. You set the task and sign off on the result.

If so far you’ve only chatted with AI, try giving it a task involving files: you’ll see the difference the first time. Why the same request gives different results is covered in detail in the article “Why ‘I tried AI and it all came out badly’ isn’t about AI”.

Claude Code plans and installation

Plan. Claude Code is included in all paid Claude plans: Pro at $20 billed monthly or $200 for a year upfront (about $17 a month), Max from $100 a month, and the team plans, Team and Enterprise. Prices are before tax. Claude Code isn’t on the free plan. There’s also a route without a subscription, pay-as-you-go through the Claude Console, but for a beginner a subscription is simpler and more predictable.

Where to run it. For a beginner the easiest start is the Claude desktop app: it has a Code tab, which is the same agent but with familiar windows instead of a command line. Other options are the terminal and the extensions for VS Code and JetBrains. The browser at claude.ai/code and the mobile app start a cloud session: it runs on Anthropic’s servers against a GitHub repository, not your computer’s folders, and keeps working when your laptop is closed. For the steps below, where the agent works through your files, you need a local session.

Installing in the terminal, if that’s what you want. On Windows, in PowerShell:

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

On macOS and Linux:

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

Then, in the folder with your files, type claude and sign in with your Claude account. This installation updates itself.

How to use Claude Code: nine steps

Each step is one action and a ready-made phrase you can say to the agent.

Step 1. Give it a small task and watch how it works

Open Claude Code and give it any simple task: rename files by a pattern, build a table from several documents, draft a letter based on an email thread. The point of this step is to see with your own eyes how the agent works through the steps on its own.

A good habit from day one is to ask for a plan first. Claude Code has a plan mode for this. In the terminal, press Shift+Tab until plan mode on appears at the bottom; in the desktop app, choose it in the mode switcher. In this mode the agent studies the files and prepares a plan without touching the project files themselves. Fixing a plan is cheaper than redoing the result.

Draw up a plan for how you’ll do this task, show it to me, and after I say “yes”, do it yourself.

If the agent heads the wrong way, press Esc and it will stop. The /rewind command, or a double Esc with an empty input field, opens the rewind menu: you pick a point and choose what to restore (the conversation, the files or both). It’s not a substitute for a backup: only the edits the agent made with its own editing tools are rolled back.

Step 2. Put the files in a folder and ask for CLAUDE.md

Don’t upload documents to the window one by one. You don’t describe the company out loud to a new employee on their first day; you give them a folder of documents. Do the same. Create a folder and put copies of everything relevant into it: price lists, contracts, correspondence, descriptions of services, spreadsheets. No need to sort them in advance, put them in as they are. A small folder is enough for the first time.

Open this folder in Claude Code and ask it to go through it. The agent will sort the files by topic, build a knowledge base and ask questions along the way. The files give it facts about your business that aren’t on the internet. Ask it to rely on them, cite the specific document and say plainly what data is missing: that way it makes things up less.

Study this folder, sort the files by topic, build a knowledge base and write CLAUDE.md: who I am, what I do, how to work with me.

CLAUDE.md is a memo the agent reads at the start of every session. You don’t have to write it by hand: the /init command studies the folder and creates a starter memo, and the agent adds information about you and the working rules from your answers. Read the memo and fix whatever is wrong. Anthropic advises keeping it short, no more than 200 lines. For each line, ask yourself: if I remove it, will the agent start making mistakes? If not, the line is unnecessary.

A memo in the project folder applies only there. General rules for all projects go in ~/.claude/CLAUDE.md in the user’s home folder. On top of that, Claude Code keeps its own automatic memory by default. You can view and turn it off with the /memory command.

Step 3. Talk, don’t type

The agent needs context. When you type, you start saving words, but it needs the details: who the client is, what you’re worried about, how it was done before. Speaking is faster and more detailed, and the agent will work out slips of the tongue from context.

In the terminal version dictation is built in: the /voice command turns it on, then you hold the space bar and talk. Russian is supported: dictation takes its language from the language setting in /config, and if none is set, it listens in English. The audio is processed on Anthropic’s servers, and dictation doesn’t use up your plan’s limits. For the desktop app and any other program, Handy will do: a free, open-source program that recognises speech right on your computer, with no internet. It runs on Windows, macOS and Linux.

I’m going to describe the task now. Don’t skip anything, and ask clarifying questions if something is unclear.

Step 4. Hand over one boring task in full

Don’t start with “automate the whole business”. Start with what annoys you every week: pick a routine that takes hours and hand it over in full.

Don’t expect perfection the first time. Let the agent put together a draft the way it sees it, and you correct it. That’s quicker than spending a long time thinking the brief through in advance. At first, check every stage, and automate once the process is running smoothly. This is how, for example, a legal agent grows: at first it only inserts client details into a contract template, then it drafts contracts from a template, and over time it checks clients’ contracts for risks.

There’s no such thing as an error-free agent; the final word is yours. So start with tasks where the result is easy to check. Anthropic considers this one of the most useful techniques for working with an agent: give it a way to check itself, such as a sample, a criterion or a table to compare against.

Here’s a task I repeat every week, and last time’s result as a sample. Make a draft and I’ll correct it.

Step 5. Claude Code skills: save a way of working that worked

It didn’t work the first time? That’s normal, it usually takes a few attempts. When you’re happy with the result, tell the agent to save the way of working as a skill.

A skill is an instruction in its own folder: exactly how you like this work done. It’s like with a new employee: you show them once how to put together a sales proposal, and from then on they follow the pattern and all you do is check. Technically, a skill is a SKILL.md file, and personal skills live in ~/.claude/skills/. The agent’s context permanently holds only the names and short descriptions of skills, and the full text is loaded when a task calls for it. That’s more economical than writing every rule into one memo, but the descriptions take up space too, so it’s better to delete skills you don’t need.

Remember this as a skill, so that next time you do it this way straight away.

You don’t have to write skills from scratch. Anthropic has an official skill-creator plugin that helps you create and test them. Ready-made skills can be found in the plugin catalogue (the /plugin command) and in open collections.

Someone else’s skill or plugin is someone else’s code on your computer. Anthropic’s documentation warns plainly that an installed plugin can run any code. Before installing, read what you’re installing. I covered how to do that in the article on how to check third-party skills and plugins. It’s about Hermes Agent, but Claude Code uses the same skill format, and the same checking rules apply.

Step 6. Connect your services through MCP

The agent should work where you work: in your CRM, spreadsheets, email, messenger. That’s what MCP is for, an open standard for connecting AI to external services. Essentially, it’s a plug adaptor: one end fits the agent, the other fits your service. Through it the agent can, for example, find a client in the CRM or read rows in a spreadsheet. Connectors linked to your claude.ai account are picked up by Claude Code automatically; others are added with the claude mcp add command.

You don’t need to understand this in advance: ask the agent to walk you through the steps. Stuck? Take a screenshot and send it to the agent. If there’s no ready-made connector, most services have an API, and the agent will build the connection itself from the documentation. Your part is to get the key and put it where it tells you. Don’t show the key itself to anyone, and don’t paste it into the chat.

Make the first connection read-only: let it look but not change anything. And connect only services you trust. Anthropic doesn’t vet third-party MCP servers, and someone else’s instructions can be slipped to the agent through the data a service brings in.

Connect my CRM through MCP. Explain step by step, without technical terms, where to get the key, and connect it read-only.

Step 7. Give the agent a role: one role, one window

The main question beginners ask: “Where is this agent, how do I launch it?” People look for a separate program, but an agent is a set of four parts:

• memory: the folder with files and CLAUDE.md;
• abilities: skills;
• connections: which services it can reach;
• role: who it is (a sales manager, an analyst, a lawyer).

Just like an employee who knows the company, has the skills, has access to the systems they need and holds a position. The simplest way to assemble an agent is right inside one session: you give it the folder, skills and connections and write the role. That window is the agent. Need a second one? Open a new session, give it a different role and rename it.

A role matters more than it seems. An agent that knows it’s an analyst looks for numbers and asks questions about the data instead of writing nice generalities. You can ask the agent itself to suggest which roles you need.

You’re my analyst. Here’s what you do, and here’s what you don’t do. Use my folder, skills and connections.

For big tasks Claude Code can launch subagents, helpers with their own separate task context. For example, one goes through a hundred files and returns only the conclusion to the main agent. That way the main conversation doesn’t get cluttered.

Step 8. Keep costs under control

The model for the task. By default, paid plans use Opus 5.5, a strong model for complex work. Sonnet 5.5 is cheaper and, according to Anthropic, handles most tasks well. Haiku 5.5 is for simple correspondence and routine. The most powerful, Fable 5.1, is for the heaviest and longest tasks, but it’s billed differently: on some plans it’s charged on top of the subscription, and on Max it has a separate weekly limit. Check the terms before choosing. You switch models with the /model command. The effort level (/effort) on the 5.5 models is medium by default, and for most tasks Anthropic advises leaving it alone.

If you want to try a different model in the familiar interface, there’s a separate breakdown: how to connect GPT to Claude Code through the local CCR proxy. It also covers the limitations of a ChatGPT subscription and the options for calling Codex through the CLI or a plugin. That’s a separate provider setup, not the usual choice of Sonnet or Opus in the menu.

Limits. On a subscription, usage is counted in five-hour windows, and paid plans also have a weekly limit. The limits are shared between the Claude chat and Claude Code. /usage shows how much you’ve used. If you hit the limit, you can wait for the window to reset or turn on extra usage. It’s charged at API prices: money is paid in advance, and you can set a monthly spending cap.

Why be careful here. Anthropic says that in companies the average Claude Code spend at API prices is about $13 per developer per active day and $150–250 a month. These are statistics for programmers in companies, not a forecast of your spending, but they give you the order of magnitude: the Pro subscription costs $20. With daily work, extra usage can easily cost more than the subscription itself. If you turn it on, set a cap.

One task, one session. The longer the conversation, the more each new reply costs and the more confusion there is. Between unrelated tasks, start afresh with the /clear command. If the conversation has grown, /compact will squeeze it down to the essentials, and /context shows how full the context is. If the agent has twice failed to understand a correction, don’t argue: start a new session with a more precise brief. More on where tokens leak away in the checklist for saving tokens when working with AI and in the article on context engineering.

Before you start, assess: is this task simple or complex, and which model suits it?

Step 9. Once you’re comfortable: work without you

The next level is an agent that works while you sleep and sends a report in the morning. Claude Code has several official routes for this:

• cloud sessions at claude.ai/code: they keep working when your laptop is closed, and GitHub is usually connected to them;
• scheduled tasks (the /schedule command): run in the cloud on a schedule or in response to an event, still in preview;
• Remote Control (/remote-control): control a session on your computer from your phone, but the computer has to be switched on;
• Telegram through Channels: you message a bot, and the task is carried out by a Claude Code session on your computer. The feature is in preview and works while the session is open.

A fully server-based agent in Telegram, with an orchestrator that hands out tasks to other agents, is a separate system such as Hermes Agent. I described how to install one in the guide to installing Hermes and connecting it to Telegram, and how it works under the hood in the breakdown of an autonomous agent. But take it one step at a time: first an agent on your computer, then a server.

Draw up a plan: which parts of my daily work could you do at night without me, and what needs to be set up for that.

Security: five rules for the first week

1. Know your permission mode. In recent versions, a session in the terminal and VS Code starts in auto mode: a separate classifier model checks actions on your behalf. Anthropic says honestly that this means fewer questions but is no guarantee of safety. For sensitive work, switch to Manual mode: the agent will ask before edits and commands. The mode in which the agent asks nothing (bypassPermissions) is only for isolated virtual machines.
2. Read what you approve. Responsibility for an approved command is yours.
3. Keys and passwords stay out of the chat. The agent will tell you where to put a key so you don’t have to show it in the conversation.
4. Other people’s skills, plugins and MCP servers only from those you trust. This is code that runs on your machine.
5. Keep a copy of important files. Rewind in Claude Code doesn’t replace a backup or version control.

Command cheat sheet

The commands below are for a local session: the terminal, a code editor or the desktop app in Local mode. In a cloud session in the browser and the mobile app some of them don’t work; this is marked in the table.

Frequently asked questions

Is Claude Code suitable for people who don’t program? Yes. For the tasks in this article (documents, spreadsheets, contracts, reports, going through correspondence) being able to set a task clearly and check the result is enough. You don’t need the terminal either: there’s the Code tab in the desktop app.

Can I use Claude Code for free? No. It isn’t on the free Claude plan. The simplest way in is Pro at $20 a month. Without a subscription you can pay as you go through the Claude Console, but then costs depend on the amount of work.

How is Claude Code different from the ordinary Claude chat or ChatGPT? A chat works with what you’ve uploaded to it. Claude Code works with files and programs on your computer, connects to services and takes the task through to the result on its own.

Does Claude Code work on Windows? Yes. There’s a desktop app and a one-command installation in PowerShell.

Can I use Claude Code from Thailand and from Russia? Thailand is on the official list of countries where Claude is available. Russia isn’t on that list.

You don’t need courses to get started: take one small task each day and save the instructions that work as skills; that way the agent gradually learns your work. If you want to understand how grown-up development with agents works, the next step is forty rules for an engineering approach to vibe coding.