What is Claude Code, and how is it different from Claude?
Claude Code is the version of Claude that works on files instead of in a chat window. Chat answers a question and hands you text. Claude Code opens a folder, writes files into it, runs the commands to test them and publishes the result. That difference decides which one you should open.
| What you want to do | Where it belongs |
|---|---|
| Ask a question, draft a text, read a document | Chat on claude.ai |
| Produce something that has to land as a file on your computer | Claude Code |
| A task with twenty steps that has to run start to finish | Claude Code |
| Work that comes back every week | Claude Code, with a skill attached |
| Generate an image or a video | Neither, Claude does not generate images |
The practical consequence is that Claude Code can be wrong in a way chat cannot. Chat gives you a bad paragraph and you delete it. Claude Code rewrites forty files. That is why the permission setting and version control, further down this page, matter more than any prompting trick.
How do you install Claude Code?
With one command in a terminal, or with no terminal at all if you use the desktop app. The installer is native, so you no longer need Node.js or npm for it.
On macOS, Linux and WSL:
curl -fsSL https://claude.ai/install.sh | bashOn Windows, in PowerShell:
irm https://claude.ai/install.ps1 | iexOn Windows, in CMD:
curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmdOpen a new terminal window afterwards and run claude --version. A working installation prints a version number. If your shell says the command is not found, the install directory is not on your PATH yet. Then start Claude Code inside the folder you want to work in with claude.
| Requirement | What you need |
|---|---|
| Operating system | macOS 13 or later, Windows 10 1809 or later, Ubuntu 20.04, Debian 10 or Alpine 3.19 and later |
| Hardware | 4 GB RAM or more, x64 or ARM64 |
| Shell | Bash, Zsh, PowerShell or CMD |
| Network | An internet connection, and a supported country |
| Plan | A paid plan, because the free tier does not include Claude Code |
Worth knowing
If you have never opened a terminal, do not start there. The desktop app for macOS, Windows and Linux has Claude Code built in, and you get the same tool without the install step. Move to the terminal later, when you know what you want from it.
On native Windows, installing Git for Windows is optional but recommended, because it gives Claude Code the Bash tool. Without it, Claude Code runs shell commands through PowerShell instead. WSL setups do not need it. Native installations update themselves in the background, so there is no separate update routine to remember.
What does Claude Code cost, and is there a free version?
There is no free version. Claude Code starts at the Pro plan, which is 20 dollars per month or 17 dollars per month billed annually. Max starts at 100 dollars per month and buys you more usage, not more features. Anthropic bills in local currency, so a buyer in the eurozone sees 18, 15 and 90 euro for the same three plans.
| Plan | Price per month | Claude Code | Who it is for |
|---|---|---|---|
| Free | 0 | No | Trying out chat, not Claude Code |
| Pro | 20 dollars, or 17 billed annually | Yes | Almost everyone starting out |
| Max | From 100 dollars | Yes | People who work in it most of the day |
Two things are worth saying plainly about that table. The first is that Max does not make Claude Code better, it makes it last longer, so moving up only helps once you genuinely hit the ceiling of Pro in your daily work. The second is that the subscription and the API are separate purchases. If you have an API key with credit on it, that credit does not pay for your Claude Code subscription, and the other way round.
Our view
Pricing is the only part of this subject where search demand is rising while everything else falls. That tells you where the hesitation sits: people are not asking whether the tool works, they are asking what it will do to their monthly bill. The honest answer is that the plan is the small number. The expensive habit is further down this page.
How do you use Claude Code for the first time?
By giving it a folder, not a goal. The first session goes wrong for most people because they point it at everything they own and ask for something large. Four steps avoid that.
- Create an empty folder for this one project and open it in Claude Code.
- Let Claude initialise that folder, so a CLAUDE.md file appears with what it should know about the project.
- Describe the first task in one or two sentences and let it make a plan before it touches anything.
- Read the plan, correct it, and only then let it run.
Step one sounds trivial and is not. Claude Code works inside the folder you point it at. Point it at your entire documents folder and you have given it access to everything in there, every search gets slower, and the blast radius of a mistake is your whole archive. One folder per project keeps both the context and the damage small.
Which setting costs you the most time?
The permission mode. By default Claude Code asks for confirmation before it acts, and that is where beginners lose hours without noticing: you give an instruction, you go and do something else, you come back twenty minutes later and nothing has happened because it is still waiting on your answer.
| Mode | What it does | When to use it |
|---|---|---|
| Ask every time | You confirm each action | Work on a live site or on data you cannot lose |
| Auto edit | Editing files is allowed, running commands is not | A sensible middle ground for content work |
| Allow everything | No interruptions at all | Only in a folder where nothing is irreplaceable |
| Plan first | Claude shows its plan and waits for your approval | Anything touching more than two or three files |
Plenty of tutorials tell you to allow everything immediately, because mistakes are cheap to undo and waiting is expensive. That reasoning holds in a throwaway folder. It does not hold on a live website, a folder of client files or a repository you use every day, where the one mistake that does cause damage costs more than every minute you saved.
The combination that works in practice: plan first for anything that touches several files, then let it run without interruptions. You see which files will change before anything happens, and after that it does not stop to ask.
Why do you need Git and GitHub if you do not write code?
Because otherwise you have no brake. Version control keeps every version of every file, so you can return to the state of five minutes ago. That was always useful. It becomes necessary the moment something other than you edits your files, because an agent rewrites a hundred files faster than you can read them.
It is also the honest answer to the permission question above. People who allow everything are rarely being reckless. They are working in a setup where any mistake costs one command to undo. Without version control you do not have that safety net, and the same setting means something very different.
| Action | What it does | When |
|---|---|---|
| Create a repository | Makes the place where your versions are kept | Once per project, set to private |
| Commit | Records the current state of your files with a short description | Every time something works, before you start the next thing |
| Push | Sends that recorded state to GitHub, so it lives outside your computer | After a commit, as a backup and to be able to publish |
| Revert | Puts your files back to an earlier commit | As soon as a change breaks something that used to work |
You do not have to learn those commands. You ask Claude Code to create the repository and to commit, and it does. What you do decide is the rhythm, and that is where beginners go wrong: they ask for one commit at the end of the day. Then you have a single block with eight hours of work inside it, and reverting means losing all of it. Ask for a commit every time something works, even if that is ten times a day. Every commit is a point you can go back to.
Two things never belong in a repository: your keys and passwords, and files a tool generates for itself. Keep the repository private until there is something in it you want the world to read, because a public repository is readable by everyone, including the people who scan for exposed keys.
What do you put in CLAUDE.md?
The rules you are tired of repeating, and nothing else. CLAUDE.md is the file Claude Code reads at the start of a session and keeps consulting while it works, so every line in there costs you something on every action. Three to five hard prohibitions do more than twenty friendly guidelines.
| Rule | What it prevents |
|---|---|
| Improve the skill, not just this one result | Giving the same correction a hundred times |
| Give the absolute path of whatever you created | Hunting through folders |
| Test before you answer, never report a result you did not run | An error you have to discover yourself |
| Do not agree with me, say so when my direction is wrong | Confirmation of a bad plan |
| Say it when I am making this more complicated than it needs to be | Systems nobody maintains |
| Never invent a figure, say when you do not know | A claim you have to retract later |
For anyone working with content, the last rule is the one that matters most. The second to last saves the most time. Keep the file short enough that you can read it in under a minute, and prune it whenever you add something.
When should you create a skill?
The third time. A skill is a written-down procedure that Claude Code loads when the situation matches: what it does, when it applies, and the steps it follows. Before the third time, writing it down costs more than explaining it again.
| Occasion | What to do | Why |
|---|---|---|
| First time | Just explain it | Explaining is faster than writing a procedure |
| Second time | Explain it again | Still faster than writing it down |
| Third time | Write the skill | From here the skill pays for itself |
The reason this threshold is worth respecting is that the alternative is invisible. Forty messages to get one thing right feels like work, so it does not register as waste. The next week you start from zero with the same explanation and the same corrections, and it still does not register, because each session on its own looked reasonable.
Most businesses run on something like twenty tasks that keep coming back. Write those down and you stop having individual instructions and start having a system you could hand to somebody else.
What are subagents, and why do they save money?
A subagent is a separate session that does one job and reports back a summary. It saves money because the search, the reading and the dead ends happen in its context window instead of yours, and only the conclusion comes back.
That matters more than it sounds, because the main conversation is the thing that gets expensive. Send three subagents to read three large files and your own session grows by three paragraphs instead of three files. Use them for anything that involves looking something up, comparing options or checking many items, and keep the decisions in the main session where you can see them.
How do you connect Claude Code to your own tools?
Through MCP, the open standard for giving an assistant access to an outside system. A connection gives Claude Code the ability to read and change things in a tool you already use, rather than you copying data back and forth.
The useful warning is the same one that applies to CLAUDE.md: every connection you switch on adds text to every conversation, whether you use it or not. Turn on the ones this project needs and leave the rest off. A practical example of one connection, end to end, is in the piece on connecting Claude Code to WordPress, and the broader picture is in AI automation services.
Where do you keep API keys and passwords?
In a separate file that version control ignores, and in the settings of whoever hosts your site. Never in the files themselves and never in the repository, because a key that has been pushed once is compromised even after you delete it, and rotating it is the only real fix.
Ask Claude Code to set that up on the first day rather than the day you publish. It is three minutes of work at the start and an afternoon of cleanup later.
How do you check that something works before you ship it?
By letting Claude test it rather than looking yourself. There are tools that let it drive a browser: it opens your page, views it at phone, tablet and desktop width, takes screenshots and reviews them.
That changes the loop more than it sounds. Without it, you ship something, you find the fault, and you go back. With it, the fault is found before you ever see it. Put a line in your CLAUDE.md saying nothing is delivered without a check and you stop losing that round.
How do you put a site online?
In two steps: your code to a repository, and that repository to a hosting provider. You let Claude do both, but you should know what happens so you can place the errors.
- Create a private repository. That is where your code lives, and you have a backup immediately.
- Connect that repository to your hosting provider and publish.
- Add your keys separately in the hosting provider's settings, because they were not uploaded.
- Publish again, otherwise the live site still does not know those keys.
- Test the form on the live site, not only locally.
Steps three and four are where it almost always goes wrong. The form works on your own computer, you publish, and a visitor on the live site gets an error. Not because the code is wrong, but because the key is missing.
Why does your usage run out after an hour?
Usually not because of the model you picked. Four habits drive consumption, and the biggest one is the least visible.
| Habit | What it costs you |
|---|---|
| One long conversation | Every message carries the entire history with it |
| A CLAUDE.md that keeps growing | It is read again on nearly every action |
| Every connection switched on | Each one adds text to every conversation |
| Always the heaviest model | Plan with the heavy one, execute with a lighter one |
The first is the one that catches people. A conversation gets more expensive the longer it runs, because each new message resends the whole history. The same short question costs far more after three hours than it did at the start, not because the question grew but because the history did.
So: new subject, new conversation. That is not tidiness, it is the single cheapest habit available to you. Check your usage now and then, so you know where you stand before you hit the wall.
Which best practices actually matter?
Five, and they are all about constraining the work rather than improving the prompt.
- One folder per project, never your whole drive.
- Plan first for anything that touches more than two or three files.
- A commit every time something works, not one at the end of the day.
- A new conversation per subject, because history is what you pay for.
- A skill from the third repetition onwards, never earlier and never later.
Notice what is not on that list. Nothing about magic words, nothing about prompt templates, nothing about a setting that makes it think harder. Those get discussed the most and change the least.
What I learned running this across 297 pages
Two things that were not in any of the twenty guides and videos I went through for this page, both from a session where I had Claude Code review all 297 pages of this website.
The first is about where your checking belongs. Two pages received the lowest rating and that turned out to be wrong, not because the model judged badly but because my text extraction had missed the main content on ten pages. A page that appeared to have 52 words actually had 288. The judgement was correct for the data it was given. The data was not. So check your input, not only your output.
The second is more useful. Because every rating carried a list of defects, I could count which defect came back most often. Thirty-one pages repeated the same definition word for word. That turned out to be one component printing the text three times. One repair instead of thirty-one rewrites.
That is the pattern that makes this tool worth the subscription, and it is rarely what the examples show: not that it writes faster, but that it can apply the same judgement to hundreds of items so that you can count. Counting is where you find causes instead of symptoms.
Where should you start tomorrow?
With one task you do every week and currently do by hand. Make a folder for it, put a CLAUDE.md in it with five rules, do the task once together with Claude, and turn it into a skill afterwards. That is half a day, and from the third time onwards you get it back.
Do not start by building a system that does everything. Start with one task that works, because a system you do not use is more expensive than no system at all. If you want to see this applied to a website rather than described, the Claude Code SEO masterclass walks through building and ranking one, and non-commodity content explains why scale alone does not get you anywhere.