Put the genie in the box

An AI agent, running on your own computer, set up for someone who has never opened a terminal. You paste two things, answer one question, and about an hour later you have a computer that does what you ask it to.

You do not need to be any good at this. That is the whole point.

What you need

• A computer that is yours — a Mac, or Windows 11. Not a work computer: the agent gets on with things without asking you about every step, and it can see and change anything in your home folder. That is what makes it useful, and it is why it belongs on your own machine and nobody else’s.
• A paid Claude plan. Claude Code is not usable on the free one. If you do not have one, claude.ai/upgrade — the first step will send you there anyway.
• A phone, or a second screen. Signing in to GitHub shows you a short code on one screen that you type on another.
• A password manager. Any one. There is exactly one thing in all of this that you have to save and can never get back if you lose it, and this is where it goes.
• About an hour, and a coffee. Some steps download a lot and show nothing while they do it. Every one of those tells you first.

Step 1 — in your browser

Open claude.ai — a normal chat, nothing special — and paste this in:

I have never used a terminal. Walk me through getting Claude Code onto my computer, one step at a time, and wait for me to say a step worked before you start the next one.

Ask me one question first: am I on a Mac or on Windows? Ask me nothing else — decide everything else yourself.

You are a chat in my browser, so you cannot touch my computer: in this part I type every command myself. Give me one at a time, tell me where to type it, and wait until I say it worked. (The next part is different — there, the agent on my computer does the typing.)

TELL ME WHAT IS ABOUT TO HAPPEN, BEFORE IT HAPPENS

This matters more than anything else here. Every time something is about to be slow, silent, alarming to look at, or costs money, say so in one line FIRST. Not afterwards, and not while I am already worried.

  - If it takes more than a minute, say roughly how long and tell me to go and make a coffee. A progress bar that has not moved for four minutes is where people close the window.
  - If the screen will show nothing while I type, say so before I type. Nothing on screen looks broken.
  - If something opens a window or a browser, say it is about to.
  - If it costs money, say so before I click, and say roughly how much.
  - If a command looks frightening, say what it does in one sentence before I run it.
  - If nothing appears to happen, tell me that is what success looks like here.

WINDOWS
  - I need WSL, which gives me a real Ubuntu terminal inside Windows. Say what that is in one sentence before we do it.
  - There is a way to put the agent straight onto Windows and it looks easier. Do not take it. Everything after this assumes Ubuntu.
  - Check first that this machine can do it: Windows 11, or Windows 10 build 19041 or newer.
  - In PowerShell as Administrator, run `wsl --install`, then restart the computer.
  - Before the restart, tell me it is a real restart and to save anything I have open.
  - After the restart nothing opens by itself. That is not a failure — say so, because waiting for something to appear is where people give up. Tell me to open Ubuntu from the Start menu.
  - The first time it opens it sits there setting itself up for a minute or two before it says anything. Warn me first.
  - Then it asks me to invent a username and password. Before I type: say clearly that this is NOT my Windows password and has nothing to do with my Microsoft account, and that THE SCREEN WILL SHOW NOTHING AT ALL while I type it — no dots, no stars, nothing. That is normal. It is the single place people get most stuck, and they blame themselves for it.
  - From here on, "the terminal" always means the Ubuntu window. If my prompt starts with `PS C:\` I am in the wrong one, and you should say so every time you ask me to open a terminal.
  - Have me make Ubuntu the default profile in Windows Terminal, so that every new tab I open later is an Ubuntu tab. Look up where that setting lives in the current Windows Terminal documentation rather than describing it from memory. Without it, "open a new tab" later lands me in PowerShell, where nothing we are about to set up exists.

MAC
  - Open Terminal from Spotlight (Command+Space, type "terminal"). That is the only setup I need to reach a terminal.
  - I will need Apple's Command Line Tools before long — real projects want git, the tool that keeps every version of your work, and the copy of git that ships with macOS is a stub that does nothing until Apple's tools are installed. Get it over with now rather than in the middle of something.
  - WARN ME FIRST, in about these words: this is a big download from Apple, it usually takes five to fifteen minutes depending on the connection, a window will pop up and I have to click Install, and then there is nothing to do but wait. Tell me to go and make a coffee. It does not ask for a password.
  - Then have me run `xcode-select --install`.
  - The command comes straight back while the download carries on in Apple's own window. That is the trap: it LOOKS finished when it is not. Do not go on until I tell you Apple's window says it is done, and say that to me explicitly rather than assuming I know.
  - Do not install Homebrew. Nothing in this kit uses it.

BOTH, once I have a terminal
  - Tell me plainly, before I sign in to anything, that Claude Code needs a paid Claude plan to be usable, and roughly what it costs. If I do not have one, send me to claude.ai/upgrade and wait for me. Do not skip this to be polite.
  - Warn me that the next one downloads a few hundred megabytes and takes a couple of minutes with very little on screen.
  - Install Claude Code with its own installer: curl -fsSL https://claude.ai/install.sh | bash
  - Check it worked: `claude --version` answers.
  - Now start it, and use exactly this line: claude --permission-mode auto
  - Explain that line in about this much, BEFORE I run it: setting up a computer is hundreds of small commands, and the ordinary mode asks me to approve each one. Nobody reads the two-hundredth question — they just press enter, which is worse than not being asked. So instead, Anthropic's own safety check looks at each action in the background and stops anything that does not fit what I asked for. After setup that is simply how it starts, and it is one line I can change if I ever want the asking back.
  - Tell me first that it is about to open my web browser to sign me in, so the window appearing is expected rather than something going wrong.

WHEN THAT IS DONE
  Tell me to keep that window open, and that the next thing I paste goes into Claude itself, not into this chat.

It asks you one question: Mac or Windows? Then it walks you through getting Claude Code onto your computer, one step at a time, and waits for you to say each one worked.

In this part, you type the commands. The chat in your browser cannot touch your computer, so it tells you exactly what to type and where, one line at a time. Three things it will warn you about before they happen, so you know what you are looking at:

• Windows: it installs a real Ubuntu terminal inside Windows, which needs a restart. Afterwards nothing opens by itself — you open Ubuntu from the Start menu. The first time, it asks you to invent a username and password, and the screen shows nothing at all while you type the password. That is normal. It is not your Windows password.
• Mac: Apple’s own developer tools download first — five to fifteen minutes, in a window Apple opens, and it looks finished before it is. Wait for Apple’s window to say so.
• Both: the last line starts Claude Code with a setting that lets it work without asking you about every command. It explains that line before you run it.

Step 1 ends with Claude Code open on your computer and signed in. Keep that window open. The next thing you paste goes in there, not in the browser.

Step 2 — in Claude Code

In the window step 1 left open, paste this:

You are Claude Code, running on somebody's own computer. They have never used a terminal, they are watching you work, and they are trusting you with a machine they cannot fix. Set it up for them.

SAY WHAT IS ABOUT TO HAPPEN, BEFORE IT HAPPENS

Every time something is about to be slow, silent, alarming or irreversible, say so in one line first. Not afterwards. Specifically: if it takes more than a minute, say roughly how long and tell them to go and make a coffee; if the screen will show nothing while they type, say so before they type; if something cannot be undone, say that before it happens, not after. If nothing appears to happen, tell them that is what success looks like here.

FIRST — READ WHAT YOU ARE ABOUT TO RUN, AND SAY SO

Get the kit into ~/projects/mwk-genie from github.com/matewishkey/mwk-genie — its main branch, unless I tell you a different branch or commit to use. On a Mac use the tar.gz rather than git unless Apple's Command Line Tools are definitely installed — /usr/bin/git is a stub that opens a dialog and fails.

Then open ~/projects/mwk-genie/install.sh and actually read it before running it. Check it against this list, out loud, in one short line each:

  - Does it use sudo — the thing that asks for the computer's password so a command can change anything outside their own folder — or ask for a password anywhere? It should not.
  - Does it write anywhere outside their home folder? It should not.
  - Does it download from anywhere other than github.com, raw.githubusercontent.com, mise.run, claude.ai, or the release hosts those redirect to? It should not.
  - Does it delete anything? It should not.

Then tell them, in one sentence and no jargon, what it is about to do. If any answer above comes out wrong, stop and tell them what you found instead of running it. This is not a formality — they cannot read it themselves, and being the person who checks is most of what you are for here.

THEN — RUN IT

Tell them first: this takes about two to five minutes, it downloads a fair amount, most of it will look like nothing is happening, and it will not ask them for anything. Tell them to leave it alone and go and make a coffee.

Then run install.sh and let it finish. It asks nothing and needs no password. It installs a tool manager, six pinned tools, Claude Code if it is missing, and their settings.

THEN — FIX YOUR OWN PATH BEFORE YOU CHECK ANYTHING

Your shell was started before any of this existed, so `mwk` is not on your PATH and will be "command not found" even though it installed perfectly. Run `. ~/.mwk-shell.sh` once, in this session, before the checks below. A new window that they open gets it automatically; you do not.

THEN — PROVE IT WORKED, IN THEIR WORDS

  - `mwk` answers with its three commands
  - the bar at the bottom of this window shows the model, the folder and a percentage — if it does not, say so; it is the thing that will tell them when to start a fresh conversation
  - http://127.0.0.1:29200/ answers, and it is their howto — open it for them and tell them to bookmark it. A new terminal window starts it if it is not running
  - a NEW terminal window knows `mwk` — this one matters most, because a shell only reads its settings when it starts, so the shortcut does not exist in the window you are sitting in
  - ~/.claude/CLAUDE.md starts with the line "# How we work". That file is written once and never overwritten, so if they had one already, theirs was kept and none of the rules you are about to follow are in force. If the first line is anything else, say so plainly, show them both, and ask whether to merge — do not overwrite it yourself.

THEN — TWO THINGS THAT MAKE YOU BETTER AT THIS

Install one plugin and one documentation source, and nothing else. Every extra thing you add costs tokens in every session forever, so this list is short on purpose.

  - claude plugin marketplace add anthropics/claude-code — then: claude plugin install frontend-design@claude-code-plugins — so the first thing they build with a screen does not look like a template. Both lines, in that order: on a machine that has never had a plugin, no marketplace is registered, and the install alone fails with "not found in marketplace".
  - claude mcp add --transport http --scope user context7 https://mcp.context7.com/mcp — up-to-date documentation for whatever library they end up using. The --scope user matters: without it, it is registered only for the folder you are standing in, and the next thing you do is move them to a different one.

Check each worked before saying it did. If either fails, say so and carry on — neither is load-bearing.

THEN — CONNECT THEIR ACCOUNTS

Run `/mwk-onboard`. It walks them through the three accounts everything else leans on — GitHub first, because that is where their work gets saved and it has to exist before the first project; then Cloudflare and Replicate — puts each key in their store from a second tab, and proves each one answers. Do not do any of it from memory and do not do the GitHub sign-in yourself beforehand: the skill has the device-code flow, the warning that goes before it, and the checks. If they do not have a GitHub account, it walks them through making one — it is free, and without it nothing they build can be saved anywhere but this computer.

THEN — THREE SMALL THINGS BEFORE THEY GO ANYWHERE

Tell them the one thing they type: `claude` starts you, from any folder. Everything else they ask you for.

Put a shortcut to `~/projects` and to `~/mwk-work` on their Desktop — on Windows those folders are inside Ubuntu, so the shortcut points at the `\\wsl$` address — and say that is how they reach their files without a terminal.

Point out that their prompt now shows the folder they are in, and a `*` when there is work they have not saved — `/mwk-save` clears it. And that all of this comes back off the computer whenever they want by asking you, which is worth knowing before they wonder.

THEN — GIVE THEM SOMEWHERE TO WORK, AND STOP

Run `/mwk-new`. It asks them the one question — what do they want their computer to do — and makes the folder the right shape, turns on save points, puts a private copy on GitHub, and moves them into it in a NEW window. Do not build the folder by hand: the shape is the skill's job, and it has to match every project they will ever have. When that skill says it is finished in this window, so are you.

HOW TO WORK, ALL THE WAY THROUGH

  - Do it for them. Never hand them a command unless only they can run it.
  - If something genuinely needs their own keyboard — a key — say so in one line: the command, and why it is theirs to run. Never tell them to quit you; a second tab is the answer.
  - The first time they run `mwk add`, warn them BEFORE they start: it makes them a key, and it will tell them to copy one line from a file into their password manager. That line is the only way into their keys — on this computer and on the next one — and nobody can make it again. Say that before it happens rather than in the sentence afterwards. And you never look at that file yourself.
  - One line about what a step is for, then do it, then show them it worked. Not a report.
  - No jargon. If a word needs explaining, it was the wrong word.
  - If something fails, say so plainly and say what you are trying next. Never announce success you have not checked.

From here on, the agent types and you watch. It will:

1. Read the installer out loud before running it — does it ask for a password, does it touch anything outside your home folder, where does it download from, does it delete anything — and tell you the answers. You cannot read it; it can. “Trust us” is not an answer, so this is the answer instead.
2. Run it. Two to five minutes, mostly silent. Go and get the coffee.
3. Prove it worked, in plain words, rather than just saying it did.
4. Connect your accounts — GitHub (where your work is saved), Cloudflare (websites, domains, email at your own domain) and Replicate (AI models). Each one is proved to answer, not assumed. When it needs a key from you, you type it in a second terminal tab — never in the chat, so it never ends up in a transcript.
5. Make you a key of your own, the first time you store one. It will tell you to copy one line into your password manager, as an entry called mwk key. Do that, once. It is the only way into your keys — on this computer and on the next one — and nobody can make it again. The agent never sees it.
6. Make your first project — a folder with the right shape, saving turned on, a private copy on GitHub — and move you into it in a fresh window.

What you end up with

claude	starts the agent, from any folder. It gets on with the work; Anthropic’s own safety check runs in the background instead of asking you
the bar at the bottom	which model, which folder, and how full its memory is — so a slow, forgetful agent is a number you can see
http://127.0.0.1:29200/	your page — bookmark it. How all of this works, in plain English, and under it every page the agent writes for you
~/projects/<your thing>/	your work: input/ for things you drop in, archive/ for what has been dealt with
~/projects/keys/	your keys, encrypted, in a repo of their own, pushed to a private copy on GitHub
mwk add NAME	puts a key in that store — you type it, so it never goes through the chat

And eight things you can ask for by name, or just say in normal words: /mwk-wish turns an idea into a researched answer and a first version, /mwk-onboard connects your accounts, /mwk-new starts a project, /mwk-save saves and pushes it and tidies up, /mwk-learn adds to your record of what you have learnt, /mwk-review is a second opinion, /mwk-tasks is what is open across your projects, /mwk-bug reports anything in the kit that is broken.

Everything else — seeing which keys you have, a new computer, changing a key, taking it all off again — you ask the agent. There is no menu and no long list of commands on purpose.

When something goes wrong

• Paste the error back to it. Whatever red text you got, paste it in and let it work it out. This is the single most useful habit you can have.
• Take a screenshot and paste that in. It can see pictures.
• Ask what a command does before you say yes. That question is never annoying.
• “You have reached your limit” is not a fault. Every plan has a ceiling on how much you can use in a stretch, and a long session can reach it. It says when it is back. Nothing is lost, and you carry on then.
• If the genie itself is broken, say “report this bug” — it writes the report, shows you, and files it only if you say yes.

Changing your mind

“Ask me before every command from now on” turns the asking back on — one line in a settings file, and the agent changes it. “Take all of this off my computer” does what it says: your key goes to the trash, not the bin, and your work is not touched.