Learn AI: where to start and how to get good at it
This is a course, not a list of links. It goes in order, it assumes you know nothing, and it ends with you using AI for something specific that you actually do.
Most AI guidance is written for people who already know what they want. This one is written for the far more common position: you can see that this stuff matters, you have opened a chat box and typed something, the answer was fine, and you have no idea what to do next.
There are two halves. The foundations are eight short lessons that take about an hour altogether and never need reading again. The tracks are what to do with it — pick the one that matches your work and follow it from "I have never done this" to "this is part of how I work now."
---
Before anything else: the one habit
If you take nothing else from this page, take this.
Use it for something you already know how to check. Not something you are curious about — something you could grade. Your own field, your own work, a question where you would spot a wrong answer instantly.
Everybody's first instinct is the opposite. You ask it something you do not know, the answer sounds authoritative, and you have no way to tell whether it is any good. Do that for a month and you either trust it too much or dismiss it entirely, and both are wrong.
Grade it on ten things you can check, and by the end you will have a real sense of where it is strong and where it quietly falls apart. That sense is the actual skill. Everything below is in service of it.
---
Part one: the foundations
Eight lessons, in order, roughly an hour in total. Each is a few minutes and each one stands on the one before it.
- What AI actually is — a very good guesser of what comes next. Almost every odd thing these tools do follows from that one fact, so this is the lesson that makes the other seven short.
- Who makes the models — five names worth knowing, what each is good at, and where to check, because it changes monthly.
- Choosing a model for the job — two questions settle it: does being wrong cost me anything, and does this text need to stay private.
- Asking for what you actually want — four things a good request has. Most disappointing answers are missing one of them, and it is usually the same one.
- What to be careful about — it invents things, your text goes somewhere, it agrees with you too readily, and it costs money.
- Running AI on your own machine — free, private, works offline, a step behind the best. How to actually do it.
- Letting AI use tools — the jump from answering to doing, and the one habit that keeps it safe.
- Where to go next — what the basics were for, and the handful of places worth following once they are comfortable.
You do not have to finish these before starting a track. Read one and two, then jump to whichever track is yours and come back for the rest when a lesson stops being abstract.
---
Part two: pick a track
A track is a use case with a path through it. Each one starts at "I have never done this", says what to try first, what usually goes wrong, and what "good" looks like once you are past the beginner stage.
Writing and thinking — drafting, editing, summarising, and the difference between using AI to write for you and using it to think faster. The track most people should start with, whatever their job.
Coding — from asking about an error message, through generating whole files, to running an agent in your own repository. The longest track and the one where the tooling moves fastest.
Planning and running projects — turning a vague intention into a plan, keeping a plan honest as it changes, and where AI genuinely helps versus where it just makes a longer document.
Research and learning — using AI to get up to speed on something unfamiliar without absorbing its confident mistakes along the way.
Images, audio and video — the creative tools, what they are actually good for now, and the editing work that used to take a day.
Everyday admin — email, documents, spreadsheets, the small repeated jobs. The least glamorous track and probably the one that gives you the most hours back.
Building your own tools — the step after all of the above: when a prompt you keep retyping should become something you can press. This is what AIOS is, and it is closer than most people think.
---
How to actually get good at this
Four things, in the order they matter.
Use it every day for two weeks on real work. Not experiments. Real things you were going to do anyway. Two weeks is roughly how long it takes to stop thinking about the tool and start thinking about the job.
Keep the ones that worked. When a request produces something genuinely good, save it. The exact wording, not a summary of it. Almost nobody does this, and it is the single biggest difference between somebody who is fast with these tools and somebody who is not. A prompt you have refined over six months is a real asset; a prompt you retype from scratch every time is a chore.
Notice where it fails and remember the shape. It is not randomly unreliable. It has characteristic weaknesses — recent facts, arithmetic, anything where being confidently plausible is easy and being right is hard, anything about your own specific situation that it was never told. Once you can predict where the wheels come off, you stop being surprised and start checking the right things.
Move up one rung when the current one gets boring. Chat, then chat with your own documents in it, then a tool that can act, then something you built. Each rung is a real step and each one is only worth taking when the one below it feels slow.
---
What this will not do
Three honest limits, because the marketing will not tell you.
It will not make you good at something you are bad at. It makes a competent person faster and a beginner more confident, which are not the same thing. If you cannot tell good work from bad in a field, AI output in that field will look great to you and you will have no way to know otherwise.
It will not remember your context unless something puts it there. Every conversation starts cold. When a tool seems to know you, something is quietly re-sending your history — and what it sends is worth understanding, because that is the whole game.
It will not stay still. Anything you read about which model is best is out of date within a couple of months. The foundations do not move; the specifics move constantly. That is why lesson two ends by telling you where to check rather than what the answer is.
---
Where this comes from
These are the notes I wrote while working this out for myself, cleaned up and put in order. The foundations are also built into AIOS as notes you own and can edit — the same text, in your own notebook, searchable alongside everything else you write.
If something here is wrong or out of date, I would rather know. And if your use case is not on the list, say so — the tracks come from real jobs people are actually trying to do, and the ones worth building tools for are the ones several people describe the same way. The contact form reaches me.
---
Start with What AI actually is if you are new to this, or go straight to writing and thinking if you would rather learn it by doing something useful.