Put the genie in the box

Everything else on this site assumes you already have an AI mate: an agent that lives in a window on your computer and can actually do things — open your files, run things, change what needs changing. I call it the genie on this page, because granting wishes is the job.

This is how you get one. It is the first thing I would sit down and do with you, and for most people it is the only hard part. After this, the rest is just asking.

Nobody is going to make you a developer. You will not be typing code. You will be reading what the genie proposes and saying yes or no — which is the actual skill, and you already have it.

What this thing actually is

You have probably used an AI in a browser tab. You type, it types back, and that is the end of it. If it says “open that file and change line 12”, you have to go and do it.

This is the other kind. It runs on your computer, in a window called the terminal, and it can do the doing:

• Read your files. Actual files, on your actual disk — the ones you can see in Finder or Explorer.
• Write and change files. Make a new one, edit an old one, put things in folders.
• Run things. Install software, start programs, take a backup.
• Ask you first. Before anything that changes something, it stops and asks. You type yes or no. That pause is the entire safety system, and it is why the rest of this page spends so much time on it.

The name for this is an agent. The window it runs in is the terminal — the black-or-white box where you type instructions to your computer instead of clicking them. You do not need to learn it. You need to be able to look at it without your stomach dropping, which takes about twenty minutes.

Two honest sentences before you spend an hour on this. It can genuinely delete or overwrite something you cared about, the same as you can — this is not a toy that is protected from itself. And it can be wrong while sounding completely certain, which is a thing AIs do (hallucination, if you want the word for it). Everything below is built around those two facts rather than around pretending they aren’t true.

Before you start

• A laptop or desktop. Not a phone or a tablet. On Windows you set up WSL — a proper Linux terminal that runs inside Windows — and the agent walks you through it. On a Mac or Linux the terminal is already there.
• An AI subscription, probably paid. Check what yours charges before you start rather than after. See the honest bits at the bottom.
• About an hour, and a cup of tea. It is mostly waiting for downloads.

What you end up with

1. A terminal you are not frightened of.
2. An AI agent running in it, logged in.
3. It set up for a beginner — explaining itself, doing the typing for you, and showing you a little bar at the bottom that tells you how full its memory is.
4. A free GitHub account, and one folder on your computer that is your workshop. GitHub is the free service that keeps a copy of your work and remembers every version of it.
5. Proof it works: you ask it to look around, and it tells you what it sees.

There are two prompts below and they go in order. The first one gets the genie installed. The second one sits it down and teaches it how to talk to you. Most people skip the second sort of thing and then wonder why the experience feels like being shouted at in a language they don’t speak.

Step 1 — the prompt that gets it running

A prompt is just typed English. There is nothing to memorise. You copy the box below with the copy button in its corner, paste it into the agent once it is running, and press enter.

It is written to make the genie ask about you first — how much help you want, what style, whether you want the reasoning — and then match that for the rest of the session.

You are guiding me — a regular person, not a developer —
through setting up an AI agent on my own computer for the
first time. This is step 1 of the "Put the genie in the box"
project from matewishkey.com.

FIRST, ask me these three questions and wait for my answers.
Then match them for the whole session, and remember them:
  1. How much help do you want?
     (a) hold my hand, explain every step
     (b) normal, step by step
     (c) I'm comfortable, go a bit quicker
  2. What style?
     (a) plain and friendly (b) technical (c) just commands
  3. Do you want the "why" behind each step, or just steps?

THEN, before we do anything:
  - Tell me in plain English what we're about to do and how
    long it takes.
  - Tell me what it will cost me, if anything, plainly
    rather than buried.
  - Wait for my "let's go."

WHAT WE'RE DOING:
  1. Make sure I have a working terminal. On Windows, set up
     WSL and explain what it is.
  2. Install a CLI AI agent using its OFFICIAL install
     instructions — point me at the real documentation,
     don't recite an install command from memory.
  3. Help me log in. I do the logging in.
  4. Get me a free GitHub account. I create it myself, in my
     browser.
  5. Finish by asking me to give you a small real task, so I
     can see it work.

RULES for the whole session:
  - One step at a time. Before each command tell me: what it
    changes, whether it needs admin access, whether it needs
    the internet, and how we'll know it worked. Wait for my
    OK.
  - Keep every line you print under 60 characters wide. My
    terminal window is small, and long lines wrap into
    unreadable soup.
  - Anything I have to copy goes on a line of its own, with
    a blank line above and below it, and nothing else on
    that line. One command per line. Never bury a command in
    the middle of a paragraph.
  - NEVER ask me for a password, a 2FA code, a recovery
    phrase or an API token. I handle every login myself, in
    my own browser. If you ever ask me for one of those, I
    will stop.
  - When I type a password into the terminal, NOTHING WILL
    APPEAR — no dots, no stars. Warn me before that happens
    the first time, or I will think it's broken.
  - Keep everything on my machine unless we genuinely need
    otherwise — and if we do, tell me why before, not after.
  - If I get stuck, slow down and explain it a different
    way. Meet me where I am.

When we're done, tell me to come back to
matewishkey.com/projects/put-the-genie-in-the-box and paste
you the step 2 prompt, which sets up your settings.

Start now by asking me the three questions.

Step 2 — the prompt that sets it up for a beginner

Do this one straight after, in the same window.

Out of the box, an agent is configured for the people who wrote it: developers, who like short answers and know what everything means. This prompt changes four things that make an enormous difference if you are not one of those people. Here is what each one is and why it matters, because you should know what you are agreeing to before you paste it.

It teaches the genie how to talk to you. Short lines that fit your window. Anything you have to copy sitting on a line of its own with space around it, so you can see where it starts and stops — a command buried in a paragraph is how people end up pasting half of it. And a standing instruction to run things itself rather than hand you homework, because you did not sign up to be its typist.

It sets up a shortcut to start it — the safer one. There is a well-known shortcut, often saved under the name ccc, that starts the agent with all of its asking switched off. It never stops to check with you again. People use it because being asked forty times an hour is genuinely annoying, and on a fresh machine that is a bad trade: the asking is the only thing standing between “tidy my folder” and a folder that is gone. Most agents now have a middle setting — Claude Code calls it auto mode — where routine things just happen but a separate check still stops the big ones. It is not a guarantee and this prompt does not pretend it is one; it is a much better default than nothing while you are learning. You get both shortcuts set up, told plainly which is which, and ccc is not the one it points you at.

It puts a little bar at the bottom of the window showing how full its memory is. An AI can only hold so much of a conversation at once — that is its context. When it fills up, the older parts get squeezed or dropped, and the genie starts forgetting things you told it an hour ago. Without the bar, that just feels like the AI getting stupid for no reason. With it, you can see it coming and start a fresh conversation before it happens.

It gets GitHub working and gives you a workshop. One folder on your computer where your projects live, connected to your free GitHub account. GitHub is a save-point machine: every version of your work is kept, so “try it and see” stops being frightening. The genie will start in that folder from now on, which means it is pointed at your work instead of at your whole computer.

This is step 2 of "Put the genie in the box" from
matewishkey.com. Step 1 got you installed and logged in.
Now set yourself up for the person you're talking to: me,
and I am NOT a developer.

Work through the five parts below in order. Before each
part, tell me in one plain sentence what it is for. After
each part, show me how to check it worked.

Do NOT recite settings, file paths, command names or flags
from memory. Check your own official documentation first,
and tell me if what I'm asking for isn't available in the
tool I'm actually running — a made-up setting that looks
right is worse than "this one doesn't exist."

PART 0 — HOW YOU TALK TO ME, FROM NOW ON
  - Assume I have never used a terminal before today.
  - No jargon. If you have to use a word like "repository",
    say what it means in the same sentence, once.
  - Keep every line you print under 60 characters. My
    terminal window is small and long lines wrap into soup.
  - Anything I have to copy goes on a line of its own, with
    a blank line above and below it, and nothing else on
    that line. One command per line. Never put a command in
    the middle of a sentence.
  - Better still, don't make me copy anything. If you can
    run it yourself, offer to run it: say what it does and
    what it changes, then do it once I say yes. I would much
    rather approve than type.
  - After anything that changes something, tell me how to
    check it worked and how to undo it.
  - Write these instructions into your own project or user
    memory file, so you still behave this way tomorrow in a
    new conversation. Show me the file when you're done.

PART 1 — A SHORTCUT TO START YOU, AND WHAT IT COSTS
  Explain to me, in plain words, what your permission
  prompts are and what happens if I turn them off.

  Then set up TWO shortcuts I can type to start you:
    - an everyday one that starts you in your safest
      practical middle setting — the one where routine
      actions run without pestering me but a separate
      safety check still reviews the risky ones. In Claude
      Code this is "auto mode"; if I'm running something
      else, find its nearest equivalent and tell me what
      it's called and what it does not cover.
    - a second one, clearly labelled as the dangerous one,
      that skips every check. Tell me the exact situations
      where that is a reasonable thing to use, and be
      honest that on my normal computer this is usually not
      one of them.

  Tell me which one you recommend I type by default, and
  why. Be straight with me: "fewer questions" is not the
  same as "safe", and I want to know the difference now
  rather than after something is gone.

PART 2 — SHOW ME HOW FULL YOUR MEMORY IS
  Explain what your context window is: how much of our
  conversation you can hold at once, and what actually
  happens when it fills up.

  Then set up a status bar at the bottom of the window
  showing at least:
    - how much of your context is used, as a percentage
    - which model I'm talking to
    - which folder we're working in

  Use your tool's own supported way of doing this rather
  than hand-writing a config file if a supported way
  exists. Then tell me the two things I should do when that
  number gets high, and roughly what number counts as high.

PART 3 — GITHUB, AND A FOLDER THAT IS MINE
  Explain what GitHub is to someone who has never used it:
  what it stores, who can see it, and why "every version is
  kept" is the point.

  Then:
    - Check whether I already have a GitHub account. If not,
      walk me through making a free one. I do the signing
      up, in my own browser. You never ask me for the
      password.
    - Get me signed in from this computer using the official
      command-line tool for it, so you can do GitHub things
      on my behalf later without me pasting anything secret.
    - Set my name and email for save points, and explain
      that both are visible to anyone who can see the
      project.
    - Make me one folder to keep projects in, somewhere
      sensible for my operating system, and tell me its
      exact full path. This is my workshop.
    - Set up a first project folder inside it, turn it into
      a repository, make the first save point, and put it on
      GitHub as a PRIVATE project. Tell me plainly how I'd
      make it public later, and what that would mean.

PART 4 — START IN MY WORKSHOP, NOT IN THE WHOLE COMPUTER
  Explain what a "working directory" is: the folder you're
  pointed at, which is the folder you can see and change.

  Set things up so that when I start you with my everyday
  shortcut, you start inside my project folder from PART 3
  rather than wherever the terminal happens to open. Show me
  how to check which folder you're in, and how to point you
  at a different project later.

  Then explain the one habit that matters most: make a save
  point BEFORE you let me try something, so there is always
  a way back. Show me the one command that does it.

FINALLY
  Give me a short list of what changed and where it lives,
  in plain English — one line each, including which files
  you edited on my computer.

  Then ask me for one small real thing I want done today,
  and do it with me using everything above.

Start with PART 0 and confirm you've understood it before
you touch anything.

The three things that trip everyone up

The password that shows nothing. The first time you type your computer password into a terminal, the screen does not move. No dots. No stars. Nothing. It looks broken, so people type it again, and again, and then get locked out. It is working. Type it and press enter. The prompt above tells the genie to warn you before this happens, which is most of the fix.

“Admin access” means admin access. When something asks for your password to install, it is asking to change your whole machine, not just one folder. That is normal for installing software and it is also exactly the moment to read what you are agreeing to. Ask the genie what a command does before you approve it. It will tell you. That question is never annoying.

The meter is running. Most capable agents bill by usage, not a flat fee. A long conversation with big files pasted into it costs more than a short one. This is not a reason to avoid it — it is a reason to know where the usage page is before you have a surprise. Find it on day one.

When you want to change something later

You do not need me for this, and that is the point. The settings above are not a one-time ceremony — they are a conversation you can reopen whenever something annoys you. Say what is wrong in ordinary words and ask for the change:

Two things about how you're set up.

First, the lines you print are still too wide for my
terminal window and they're wrapping. Keep everything
under 55 characters from now on, and update whatever file
you saved that instruction in so it sticks tomorrow.

Second, the status bar at the bottom — I'd like it to
show how much of your memory is left rather than how much
is used, because "82% left" is the number I actually care
about.

Tell me what you're going to change before you change it,
then show me the files you touched.

Anything on this list is fair game to ask for the same way: make the asking more frequent, make it less frequent, change what the bar shows, point yourself at a different project folder, stop using a word you keep using that means nothing to me. It is your machine and your genie.

Not using Claude?

This is the idea, not an advert. Any CLI agent that can read files and run commands works — Codex CLI, Gemini CLI, and others. Step 1 names none of them on purpose, and step 2 names one setting by its Claude Code name only because that setting is the whole point of that part; it also tells the agent to go and find its own equivalent and say so if there isn’t one. Fork the approach to whatever you already pay for.

Mate Wish Key is not affiliated with, sponsored by or endorsed by Anthropic, OpenAI, Google, GitHub or anyone else named on this page. They are named because that is what the things are called.

Why “in the box”

A genie that lives in the cloud and can only talk is a chatbot. A genie on your machine, that can open your files, is a different animal — more useful and more able to break things. Putting it in the box means both halves: getting it in, and knowing where the sides of the box are.

That is why the prompts spend their first breath asking about you, and their second telling you what they are about to touch. A genie that does things without asking is not a better genie.