Install Claude, and let it do the rest.
Nine levels, in order. The first one is the part everybody skips, and it is the reason most people give up by the weekend. The last has things getting done while you sleep. Take them at your own pace and stop wherever it stops being useful. You never write a line of code.
The nine levels
- 01Before you install. The part everybody skips.
- 02Get it running. Twenty minutes, and it works.
- 03The habits. Nothing to install, everything to gain.
- 04Your world, with edges. What it reaches, where it stops.
- 05Paying less. Where the money actually goes.
- 06Your own skills. Find them, make them, prune them.
- 07Several at work. Helpers, and a second opinion.
- 08Running without you. Scheduled, and watched.
- 09What breaks. The five that cost me time.
Before you install anything
Most people install it, try two clever things, find it impressive, and have quietly stopped by the weekend. Not because it did not work. Because they never decided what it was for.
Know what you are signing up for
This is not an AI that answers questions. It is something that opens your folders, writes in your files, runs commands on your machine, remembers how you like things done, and eventually gets on with work while you are asleep. That is a bigger thing than a chat window, and it asks for about twenty minutes of real attention before it starts paying you back.
If what you want is a better search box, you already have one and you can close this page. If you want something that takes work off your hands, keep reading.
Make the one choice that costs money
This is the only point in the whole guide where you might reach for a card, so here it is straight. Every level below can be done on the free tier. You will be slower and you will wait more often, and it genuinely works.
Every level below, on free and local models. Slower, less sharp, and it does the job.
The daily comfort. The image models come with it, which quietly replaces a couple of paid tools.
You stop watching the meter. That is all it buys, and it is worth more than it sounds.
The honest move is to take nothing today. Do level 2 on the free version and give it three days. You will know by then whether you are the kind of person who reaches for it, and if you are not, you kept your money.
Open an empty folder, not your desktop
Nobody says this and everybody regrets it. Make a brand new folder with nothing in it, and work there for the first few days. You are going to ask for things you did not quite mean and say yes to things you did not read. Much better that it happens in a room with nothing breakable in it.
Get it running
About twenty minutes. At the end, Claude works on your computer and remembers you between sessions.
Get Claude on your computer
This is the only part you do by hand. After this, Claude does the work.
- Go to claude.ai/download and get the app for your computer.
- Open the file you just downloaded and let it install.
- Launch Claude and sign in. A free account is enough.
- Open a new chat. You are ready for the next step.
Paste this and let it set everything up
Claude checks what you already have, installs what is missing, and asks before each step. Say yes, and read what it tells you.
What it just installed, and why
- Node.js
- The engine the tools run on. Nothing to learn about it, it just needs to be there.
- git
- The one that matters most. It takes a snapshot of your files, so any change can be undone. It is what lets you say yes to Claude without fear: if the result is wrong, you go back to before.
- VSCode
- A free editor. Think of it as the room where you and Claude work, with your files on one side and the conversation on the other.
- Its memory
- A MEMORY.md file and a memory folder. Without it, Claude forgets you every time you close the window and you re-explain your projects forever. With it, you pick up where you left off.
- CLAUDE.md
- Your standing rules, read at the start of every session. Things like answer in my language, ask before anything risky. You write them once instead of repeating them.
Add Codex too, if you want
Optional. Codex is OpenAI's version of the same idea, and some people keep both to compare. You do not install it yourself: paste this into the Claude you just set up, and it handles everything. Codex needs a ChatGPT subscription or a little OpenAI credit, so skip this step to stay free.
The habits that actually change things
Nothing to install here. This is what separates people who get real work out of this from people who try it twice and drop it. Each one has the vague version and the version that works, because the difference is the whole point.
Ask for the outcome, not the steps
Most people write orders. You get much further describing where you want to end up and letting it work out how. Say what the result should look like, and what it must not do.
Work in a folder, with git watching
Open a folder in VSCode and start Claude from there. It sees your files, and git records every change. That combination is what makes it safe to let it work on real things instead of copy-pasting into a chat window.
Make it prove things instead of claiming them
An assistant will happily tell you a thing is done. Ask it to show you. Run it, open the page, print the result. This one habit catches most of the mistakes before they cost you anything.
Tell it to remember, out loud
When you correct something and you would rather not repeat it next week, say so. It writes it into its memory and applies it from then on.
Keep your passwords to yourself
Never let an assistant type your passwords, your bank details or your API keys. A well set up one will tell you where to paste it and let you do it. The prompts on this page are written that way on purpose.
Your world, with edges
Up to here it only sees the folder you opened. This level widens what it can reach, then puts the walls back up on purpose.
Give it the rest of your world
On its own, Claude only sees the folder you opened. Connectors let it reach your notes, your calendar, your files, the web. This is the point where it stops being a chat window and starts being something that actually knows your situation. Add them one at a time, and only the ones you will use.
Split your work into domains
Everything in one pile is what makes an assistant vague. Separate your worlds, give each one its own folder, its own rules, its own memory. Claude behaves differently in your accounting folder than in your content folder, because you told it to.
Paying as little as possible
This is the level nobody writes, because writing it means having hit the ceiling a few times yourself. I have. Here is where the money actually goes, and how to stop feeding it.
Know where the money actually goes
What costs you is not how many questions you ask. It is how much stuff you drag into the conversation. Measured across hundreds of my own sessions, the output of commands was the heaviest thing by a long way, then reading files. The questions themselves barely registered.
So two habits pay for themselves the same day. A command that prints a lot goes into a file, and you read the first few lines of it. A big file gets read on the part you need, not from the top. And the biggest saving of all is not even on this level: it is sending helpers to read for you, which gets its own step at level 7.
Generate your images in the browser, not the terminal
The same image models cost credits from the command line and nothing at all in the web interface you already pay for. Same models, same quality. It is a billing accident, and over a year it adds up to real money. Generate in the browser, save the file, use it locally. I learned this one the expensive way.
Run the repetitive stuff on your own machine
Every question you ask uses part of your allowance, and when you run out you wait. Ollama installs a model that runs on your computer instead: slower and less clever, but free and unlimited, and your files never leave home. You keep Claude for the hard parts and hand the routine work to the local one.
Stop making it read everything
When you ask about a project, Claude opens your files to understand them, and that is what burns through your allowance fastest. A knowledge map reads the folder once and keeps a summary, so afterwards it consults the map instead of the files. Same answers, a fraction of the cost.
Have a way out before you need one
You will hit the limit, and it will be on the evening you had plans. The difference between a lost hour and a shrug is whether the way out was already set up. Have a second tool installed and signed in, or a switch to free models ready to flip. Do it on a calm day: nobody sets up a good fallback while they are locked out and annoyed.
Your own skills, and where to find them
A skill is a written procedure your assistant reads by itself when the situation matches. You do not call it up. It recognises the moment and applies it. By now you have noticed you explain the same things over and over, and this is where that stops.
Install one that already exists
There are thousands of these written by people who are better than you at their particular thing. Installing one takes about ten seconds and can replace an afternoon of explaining. Look at what exists before writing your own, because the good ones have already been argued over by people who use them every day.
Turn what you repeat into a skill
You describe once how you want your invoices sorted, or your captions written, and from then on it follows that instead of you re-explaining. The trick is to make them out of things you already do every week, not things you imagine doing. If you have explained the same thing three times, that is the one.
Do not switch them all on
Here is the trap nobody mentions. Every active skill takes up room and muddies the choice: at a hundred of them your assistant no longer knows which one applies, so it picks badly and you lose the benefit of having any. Keep a handful switched on, leave the rest sitting on disk, and go and fetch one when a job calls for it. Mine live in an indexed folder that costs nothing until I reach for it.
Put several of them to work
One assistant doing everything hits a wall on big jobs. Several, each with one job, do not.
Send helpers to read instead of reading yourself
When a job means going through twenty files, Claude can send helpers to do it. Each one reads its part on its own side and comes back with a summary, so the main conversation stays clean and you keep room to think. This is the single biggest change once your projects get real, and it is also the biggest thing you can do for your bill.
Have one plan and another one check
Nobody marks their own homework well, and a model is no different. Have one write the plan and do the work, and a second one look for what the first got wrong. You catch things you would have shipped otherwise. It costs a bit more and it is worth it on anything that matters. The pairing I actually run has its own page: four commands, and the free ChatGPT tier covers it.
Why it works is not that the second one is cleverer. It is that whoever wrote the thing is the worst placed person to see what is wrong with it. A reader who was not in the room when the decision was made catches what the author cannot.
Let it run without you
The last build step is when things happen while you are doing something else. This is also where it gets risky, so it comes near the end on purpose.
Schedule the boring parts
Anything you do on a rhythm can run on its own: a weekly summary, sorting the month's invoices, watching a page for changes. Start with something read-only, where the worst case is a wrong summary rather than a deleted file. Once you trust it, widen it.
See what it did while you were away
Something that runs on its own and reports nothing is something you will stop trusting. Give it a log and a simple page: what ran, what it did, what failed. This is what makes the difference between a setup you actually rely on and a pile of scripts you are quietly afraid of.
What I let run alone, and what I never do
- Read something and summarise it
- Yes, without thinking about it. The worst case is a summary that is wrong, and I find out by reading it.
- Move, rename and sort files
- Yes, as long as the folder is one git is watching. Then a bad night is one command to undo.
- Send anything to anybody
- No. It writes the draft, I press send. An email that leaves on its own cannot be called back.
- Delete anything
- No, ever. Things get moved to a folder called old. Deleting is a decision I keep.
- Spend money
- No. Not a subscription, not a credit, not a single API call I did not ask for.
What breaks, and why
Every guide stops at the happy path. These five are the ones that actually cost me time, so you get to skip them.
The five I hit myself
Where the time goes
- The conversation that went cold
- Leave a session sitting for an hour and the whole history has to be read again from scratch on your next message, at many times the cost. Long pauses are expensive. Either keep going or close it properly and start clean.
- Reading a file you just wrote
- If it edited the file, the edit worked. Opening it again to check costs you the whole file for nothing. I did this hundreds of times before I bothered to measure it.
- Installing something in the middle of a job
- Adding a plugin or a connector mid-session throws away everything the conversation had loaded, and you pay to rebuild all of it. Install between jobs, never during one.
- The thing that looks done
- An interface nobody opened in a real browser is a guess, not a result. This is the most common way to ship something broken while feeling productive.
- The working file that went public
- Anything sitting in the folder you publish gets published. Not the files you listed, all of them. Keep your scratch files outside that folder, because the only exclusion you can rely on is not being there in the first place.