My setup, from the first click to an assistant that runs on its own.

Chapter 01 / 09

Before you install anything

Most people install, try two clever things, find it impressive, then quit before the weekend. Not because it doesn't work. Because they never decided what it was for.

Step 1 · Know what you're getting into

Claude Code is not an AI that answers questions. It opens your folders, writes in your files, runs commands on your machine, remembers how you like to work, and keeps going while you do something else. Count on twenty minutes of attention before it starts to pay off.

You just want a better search engine? You already have one, you can close this page. You want to hand off part of your work? Keep going.

Step 2 · Make the only choice that costs money

This is the only place on this page where you might reach for your bank card. You can follow almost everything without paying: with free ChatGPT and Codex, or with a model that runs on your computer. You will go slower and wait more often. Claude Code, for its part, needs a paid plan.

Pricing changes often: look at it on each tool's plans page, the day you choose.

Step 3 · Open an empty folder, not your desktop

Create an empty folder and work inside it for the first few days. You will ask for things you didn't quite want, and say yes to things you haven't read. Better that it happens in a room where there is nothing to break.

Chapter 02 / 09

Getting it running

About twenty minutes. At the end, Claude works on your computer and remembers you from one session to the next.

Step 4 · Install Claude on your computer

This is the only part you do by hand. After that, Claude does the work.

• Go to claude.com/download and download the app for your computer.
• Open the downloaded file and let the installation finish.
• Start Claude and sign in with your paid account.
• Open the "Code" tab, choose "Local", then select your empty folder.

Step 5 · Paste this and let it install everything

Claude looks at what you already have, installs what is missing and asks before each step. Say yes, but read what it tells you.

You are my setup assistant. I have never coded and I just installed you. Configure my computer so you work well, and explain everything in simple words.

Rules:
- One step at a time. Before each installation, tell me in one sentence what it is for, then ask "shall I run it?" and wait for my go-ahead.
- Check what I already have before installing anything.
- First find out which system I have (Windows, macOS or Linux) and use the right method.
- If I have to click somewhere, tell me exactly where.
- Never type my passwords or my keys. If one is needed, tell me where to paste it and I will do it.
- At the end of each step, tell me in one line what we just gained.

Do this in order:
1. Tell me which system I have and what is already installed (Node.js, git, VS Code). Just the list, do not install anything yet.
2. Install what is missing: Node.js, git (it is what lets us undo any mistake), and VS Code (the editor where we will work).
3. Add the Claude extension in VS Code, then show me how to open a folder and start you from there.
4. Set up your memory so you remember me from one session to the next: a MEMORY.md index and a memory/ folder, one piece of information per file.
5. Write a CLAUDE.md with my default settings: answer me in my language, ask before any risky or hard-to-undo action, and make sure any text I am going to publish does not sound like a robot.
6. Have me do one real thing from start to finish: create a folder, ask you for a small change, show me exactly what changed, and show me how to undo it.

Then give me a short recap: what is installed, and the three habits to remember for tomorrow.

Step 6 · Add Codex, if you want

It's optional. Codex is OpenAI's coding agent: the same idea, from another company. You don't install it yourself: paste this prompt into the Claude you just set up. Codex works with any ChatGPT account, even a free one, with tight limits on Free and Go (OpenAI help (new tab), checked on October 6, 2026), or with an API key.

You are my setup assistant. Install OpenAI Codex on my computer. I don't code, so explain everything in simple words.

Rules: one step at a time, ask before each installation, check what I already have, find my system first, tell me exactly where to click, and never type my passwords or my keys yourself.

Be honest with me from the start: Codex works with any ChatGPT account, even a free one (tight limits on Free and Go), or with some credit on the OpenAI API. Find out what I have and take the simplest path for me.

Do this in order:
1. Tell me what I already have: Node.js, git, VS Code, and whether I have a ChatGPT account or an API key.
2. Install what is missing.
3. Install Codex.
4. Help me sign in. I type the password or the key, never you.
5. Set cautious defaults: ask me before writing or deleting files, and only work in the folders I allow.
6. Have me do one real task, then show me how to undo it if I don't like the result.

Then a short recap: how to start Codex tomorrow, and the three habits to remember.

Chapter 03 / 09

The habits that really change things

Nothing to install here. This is what separates the people who get real work out of it from those who try twice and give up. All six fit in six lines:

• Describe the result, not the steps, and say what it must not touch.
• Work in a folder tracked by Git, backed up privately on GitHub.
• Ask for proof: "run it and paste what it shows", not "does it work?".
• Have it read the current docs before it touches a technical setting.
• Tell it to remember, clearly, when you correct something.
• Keep your passwords to yourself: it tells you where to paste them, you do it.

Each habit has its example and its prompt in the guide No. 03 The 6 habits to pick up from day one.

Chapter 04 / 09

Your own world, with limits

So far, it only sees the folder that is open. This level widens what it can reach, then puts limits back on purpose.

Step 7 · Give it access to the rest of your world

Connectors give it access to your notes, your calendar, your files, the web. This is where it stops being a chat window and starts to know your situation. Add them one by one, only the ones you will use.

I want you to be able to reach beyond the folder I opened. Explain connectors to me, one by one.

1. Explain in three lines what a connector is and what it changes for me.
2. Show me the ones worth having for someone who doesn't code, and what each one lets you do.
3. Let me choose, then install only the ones I picked. Don't install a whole list.
4. Test each one in front of me right after installing it.

Before each one, tell me what it will be able to see or do, so I can decide knowingly. Never handle my passwords or my keys yourself.

Step 8 · Split your work into areas

Putting everything in the same pile makes an assistant vague. Give each area its own folder and its own rules. Claude then behaves differently in your accounting folder and in your content folder, because you told it to.

Everything I do is in the same pile, and it makes you vague. Help me split it into areas.

1. Ask me what my main areas of work are, and what lives in each one.
2. Suggest one folder per area, and tell me what goes where.
3. For each area, write a small rules file: how you should behave there, the tone, what must never be touched, what I care about.
4. Don't move anything without asking me first.

Then show me how to check that I'm in the right place before I start working.

Chapter 05 / 09

Paying as little as possible

Nobody writes this level, because you have to hit your limit several times to write it. That's my case. Here is where your quota really goes, and how to stop losing it.

Step 9 · Know where the money really goes

What costs you is not the number of questions. It is everything you let into the conversation: what commands print, the files it reads. The questions themselves weigh almost nothing.

Two habits pay off the same day. A command that prints a lot writes to a file instead, and you read the first lines. A big file is read where you need it, not from the top. The biggest saving comes at level 7: sending helpers to read for you.

I want to know what really uses up my quota, and fix the worst of it. Measure before changing anything.

1. Go through my recent sessions and tell me what filled them, by category: command outputs, file reads, my own questions, tool results. Give me percentages, not impressions.
2. Name the three biggest items in my case, with the commands or files involved.
3. For each one, give me the light version: the command that writes to a file instead of printing everything, the file read only where it matters instead of in full.
4. Write these rules down in black and white, so you apply them from now on without me asking again.

Measure first and show me the numbers. If my usage is already lean, say so instead of inventing work.

Step 10 · Generate your images in the browser

With some services, the same image models cost credits on the command line, and nothing extra in the web interface you already pay for. Before automating images, check how your service bills the two.

Step 11 · Run repetitive tasks on your machine

Every question uses part of your quota, and when it's empty you wait. Ollama installs a model that runs on your computer: slower and less clever, but free and unlimited. Keep Claude for what is hard, and give the routine to the local model.

I want part of the work to run on my machine instead of using my quota. Install a local model, and explain each step in simple words.

1. Look at my hardware and tell me honestly what size of model it can run properly. If my machine is too limited for this to be useful, say so and stop there.
2. Install Ollama, and download a model that suits my machine.
3. Test it in front of me so I can see it answer locally.
4. Explain in two lines when to use the local model and when to keep asking you.

Ask before each installation, and tell me what each step brings me.

Step 12 · Stop making it read everything

When you ask a question about a project, Claude opens your files to understand. That is what empties your quota fastest. A knowledge map reads the folder once and keeps a summary. After that, it consults the map instead of the files.

Reading my project files is what uses up my quota fastest. Set up a knowledge map so you read a summary instead of opening everything.

1. Install graphify, or the closest free equivalent if it isn't available on my system. Tell me honestly which one you use.
2. Connect it to my local model if I have one, so building the map costs nothing.
3. Build the map of one of my folders, and show me the before and after: what reading this folder used to cost, what it costs now.
4. Give me the two commands I need: one to build a map, one to query it.

Explain what you are doing at each step, and ask before installing.

Step 13 · Plan an exit before you need one

You will hit your limit, and it will be on the evening you had something else planned. The difference between losing an hour and shrugging is having prepared the exit. Do it on a calm day: nobody builds a good Plan B while stuck and annoyed.

I want to be able to keep working when I hit my limit. Set this up now, while nothing is broken.

1. Tell me my options if I run dry in the middle of a task: a second tool, free models, a local model. Be honest about what it costs and what we lose in quality.
2. Set one of them up completely, connected and tested, so it works the day I need it.
3. Give me a single command or a single click to switch, and put it somewhere I will find it when I'm annoyed.
4. Write down what to do when it happens, in five lines, assuming I have forgotten everything.

Test the switch in front of me today. A safety net nobody has tested is not a safety net.

Chapter 06 / 09

Your own skills, and where to find them

A skill is a written procedure that your assistant reads on its own when the situation matches. You don't call it: it recognizes the moment and applies it. By this stage, you have noticed that you explain the same things over and over. This is where that stops.

Step 14 · Install one that already exists

There are thousands, written by people who know a subject better than you do. Installing one takes ten seconds and sometimes replaces an afternoon of explaining. Look at what exists before writing your own: the good ones have already been proven by people who use them every day.

I want to use skills that other people have already written, instead of explaining everything myself. Help me find some, install them, and not pile them up.

1. Explain in three lines what a skill is and how it triggers without me calling it.
2. Show me where the good ones are published, and find the ones that fit what I actually do. Ask me what I do before you go looking.
3. Install two or three, and show me one running on a real task of mine.
4. Then the part nobody mentions: tell me how many I can keep active before you start choosing badly between them, and set up a place on disk for the rest.

Don't install a long list. Two I actually use are worth more than twenty that sleep.

Step 15 · Turn what you repeat into a skill

You explain once how you sort your invoices or how you write your captions, and it follows that rule instead of making you start over. Start from what you already do every week, not from what you imagine doing. If you have explained the same thing three times, that's the one.

I keep re-explaining the same things to you. Turn one of them into a reusable skill.

1. Look at what I have asked you several times and suggest the three that deserve to become a skill. Pick from what I really do, not from what sounds impressive.
2. Let me choose one, then ask me the questions you need to write it properly.
3. Write it, then read it back to me in plain words before saving it.
4. Test it right away on a real case of mine, and fix whatever came out wrong.

Keep it short. A skill that is too long ends up not being followed.

Step 16 · Don't switch them all on

The trap nobody mentions. Every active skill takes up space and blurs the choice: with a hundred active skills, your assistant no longer knows which one applies and chooses badly. Keep a few active, leave the others on disk, and go and fetch one when a task calls for it.

Chapter 07 / 09

Put several of them to work

A single assistant ends up stuck on big subjects. Several assistants, each with its own task, don't.

Step 17 · Send helpers to read for you

When a task means going through twenty files, Claude can send helpers. Each one reads its part on its own side and brings back only a summary. The main conversation stays clean, and you keep room to think. It is the best lever on your quota.

When a job means going through a lot of files, I want you to send helpers instead of reading everything in our conversation.

1. Explain in three lines what a helper is, what it sees, and what it gives back to you.
2. Take a real job of mine that touches several files and do it that way, in front of me.
3. Show me the difference: what reading everything here would have cost, what it cost this way.
4. Tell me the two or three situations where it is worth it, and the ones where it is overkill.

Step 18 · Let one plan and the other check

Nobody grades their own homework well, and a model doesn't either. Let one write the plan and do the work, then ask another to look for what the first one missed. It costs a bit more, and it is worth it for what matters. The second one is not smarter: it wasn't there when the decision was made, so it sees what the author no longer sees.

I want a second pair of eyes on everything that matters, instead of you checking your own work.

1. Take something I am about to finish. Write the plan, then do the work.
2. Then bring in a separate reviewer, and ask them to look for what is wrong, missing or risky. Not to be polite.
3. Show me what the reviewer found, and fix what deserves fixing.
4. Tell me honestly when the extra cost is worth it and when it isn't.

My duo is Claude building and Codex reviewing. How I split the work between the two: Claude and Codex together.

Chapter 08 / 09

Let it run without you

The last building step: things move forward while you do something else. It is also where the risk arrives, so I put it near the end on purpose.

Step 19 · Schedule the boring parts

Anything you do regularly can run on its own: a weekly summary, sorting the month's invoices, watching a page. Start with something read-only, where the worst case is a bad summary and not a deleted file. When you trust it, widen the scope.

I want something to run on its own, at regular intervals. Start with something harmless.

1. Ask me what I repeat at a regular pace, and pick one where the worst that can happen is a wrong summary, not lost work.
2. Set it up on that schedule on my machine, and explain in plain words what triggers it.
3. Run it once in front of me so I see the result before I trust it.
4. Show me how to turn it off, and where to look if it stops working.

Nothing that deletes, sends or publishes anything without me. Not at this stage.

Step 20 · Look at what it did while you were away

Something that runs on its own without saying anything ends up not inspiring trust. Give it a log and a simple page: what ran, what it did, what failed. That is the difference between a setup you rely on and a pile of scripts that worries you.

Things now run without me, and I want to know what happened.

1. Make everything that runs on its own write to a log: what ran, when, what it did, what failed.
2. Build me a simple page I can open to see this at a glance. No terminal.
3. Add a clear signal when something failed, so I don't have to go looking for it.
4. Show it to me with real entries in it, not an empty screen.

Keep it plain. I want to understand in five seconds, not be impressed.

Chapter 09 / 09

What breaks, and why

Every guide stops when everything goes well. These five things really cost me time. Learned by getting it wrong, so you don't have to.

Step 21 · The five I ran into myself

• The conversation that went cold. Leave a session paused too long, and the whole history has to be reread at the next message, for much more. Keep going, or close it cleanly and start a fresh session.
• Rereading a file that was just written. If it edited the file without an error, the edit worked. Reopening it to check costs you the whole file for nothing. In my audited sessions, I had done it 826 times.
• Installing something in the middle of a task. Adding a plugin or a connector mid-session throws away what the conversation had kept in memory, and you pay to reload it. Install between two tasks.
• The thing that looks finished. An interface nobody has opened in a real browser is a hypothesis, not a result.
• The working file that became public. Everything in the folder you publish will be published. Not the files you listed: all of them. Keep drafts out of that folder.

Two reflexes before publishing

The security audit. Before putting a project online, ask it for an audit of the folder: an API key visible in the code, a password forgotten in a file, a setting that is too open. It finds them in a few minutes. You, in production, find them too late.

The key that lay around. An API key pasted into a shared conversation or a published file is burned, even if you removed it since. You don't hide it: you revoke it, and create a new one from the service's dashboard.