An LLM is not a chat box that answers questions. An LLM is a million brains in one, and the right ones turn up for whatever you are doing. The course runs from your first hour to your hundredth — for writers, marketers and people who design and build.
Three tracks, six stages. The first two weeks are shared — everyone unlearns the same thing first. The tracks split at week three.
A million brains in one. The right ones turn up for your task, work faster than you can, and get better at you.
Twenty-five days in order for each of the three tracks, both week reviews and six companion pieces. Sixty-eight pages, all open.
2,995 of them, said plainly — what it means, the move it implies, one line you could type, and the trap.
Nothing is scored. Getting one wrong is the useful part.
Twenty of these. Sit with one, write down what you decide. The answer comes back three days later.
260 entries. The ideas, the file types, the acronyms — each with a handle you remember it by, before the explanation.
A hundred real jobs, sorted by day one, week one and month one. Each shows the ask, the move and what goes wrong.
Each one was paid for once — a lost day, an emptied file, a bill nobody saw coming.
The same first fortnight for everyone, then your own work. Writing, marketing, design and code — all three run to the end.
1,694 of them, each dated and kept as it happened: what went wrong working with an AI, why, and the rule that came out of it.
You arrive thinking this is a chat box that answers questions. One wrong picture holds most people back, and the problem is not ignorance — it is a wrong picture, which is harder to shift than an empty one. The first three chapters replace it, then introduce Anthropic and Claude.
Your track begins. The fastest real win is not a clever way of asking — it is to stop typing at it and point it at what you already have. Your documents, your spreadsheets, the sixty files you have been avoiding. There is a catch nobody mentions: you cannot point at what you cannot find, so this week also covers getting your own machine in order.
Now it stops being one-off asks. You write down the standard you keep repeating so you never type it again, you learn a subject you do not know yet, and you find out what to do when the answer is wrong rather than merely disappointing. The week ends on the thing most people get wrong for months: choosing what to hand over, and what to leave out.
The tracks part company here — up to now everyone has been doing the same thing. This is the writing and instructional design track, written end to end. It starts with the part people fear losing, your own voice, and ends with one real piece you made rather than an exercise. In between: turning a pile of sources into a structure, drafting, cutting, checking evidence, and building something someone can actually learn from.
Almost none of this is about how you ask. The skill is knowing where to put an instruction so it is actually followed, when to start again and what to carry over, how to have one run check another's work, and what it all costs. The gap separates people who can use an LLM from people who can run things with one.
The days for this stage are not written yet — the rules that came out of it are.
Harder questions, worked through properly. Fifty-five pairs of things people treat as one — plausibility and truth, refactoring and rewriting, a research gap and missing information — and the wrong prompt, workflow or decision that follows from mixing them up. For people who have put the hours in.
Ten shared days, then twenty-five more in whichever track is yours — writing, marketing, or design and code. Six companion articles and the reflections sit alongside. Set aside ninety minutes a day: an hour doing it, an hour thinking about it. Less than that and it will not stick.
Two things people treat as one, and the wrong prompt, workflow or decision that follows.
It re-reads the entire conversation every time you speak, and there is a ceiling on how much it can hold. Past that ceiling the oldest part falls off the back. It has not lost interest. The thread grew longer than the limit, and the beginning is what goes first.
Ask for the source, then open the source. A fabricated citation collapses as soon as it is clicked, which makes clicking the fastest test available. Where nothing can be checked, treat the answer as a draft to verify rather than a fact to repeat.
Not the public — this is not a forum post. The real question is whether your text is kept and used to improve the model, and that differs between a personal account and a company one. The setting exists and is findable. It is worth ten minutes before pasting rather than after.
Length is not counted in messages. Every turn re-sends the whole conversation, so a long thread costs more each time you speak, and an attached file is re-read on every single turn. One short chat carrying a big document can cost more than fifty ordinary ones. Starting fresh resets the bill.
There is deliberate randomness in how the next word gets chosen, so identical input does not guarantee identical output. That is a feature when writing and a nuisance when checking facts. If two answers disagree on something that matters, the disagreement is worth chasing.
Two questions, and they get collapsed into one. Detection is unreliable in both directions — it clears real machine text and accuses real people. Whether it is permitted is a rule someone has written down, or has not. Go and find the rule. Do not calibrate to the odds of being caught.
With one task you already do and dislike. Not a course, not a tour of the features. The people who get somewhere start from work they own, because they can tell whether the output is any good — and that judgement is the part no tool supplies.
Three ways into the same thing, built for different jobs. Chat is for thinking with it. The coding tool sits on your machine and works on your real files. The API is for wiring it into something that runs without you. Most people need the first two.
Less than the internet suggests. They differ at the edges — tone, file handling, behaviour over long documents — and those differences move every few months. Fluency in one transfers almost entirely to the others. Pick one, get good, stay curious rather than loyal.
Twenty-five days, and about ninety minutes a day — an hour doing it and half an hour thinking about what happened. Less than that and it will not stick, because the thinking is where the habit forms. The first two weeks are shared by everyone. From week three you follow one of three tracks, and each track is fifteen more days.
In order for the first two weeks, and it matters. Almost everything people call an AI problem is solved in those ten days, and every later day assumes them. After that, the days inside your track can be taken in any order you like — they are separate skills rather than a chain.
Pick the one whose work you actually do, not the one that sounds most advanced. Writing and instructional design covers drafting, editing, evidence and turning material into something people learn from. Marketing covers positioning, offers, numbers and briefing. Design and code covers files, bugs, review, cost and shipping. All three are written end to end.
It is free and there is no account. Nothing on these pages calls an AI service, so there is no usage to bill and nothing to sign up for. You will need access to an assistant of your own to do the exercises, and the free tier of any of the major ones is enough for the whole course.
Nothing about AI. What helps is having real work to bring — a document, a spreadsheet, a repository, a campaign. Every day asks you to point the assistant at something of your own, because the exercises are only useful when you can judge whether the answer was any good.
All of them. The course teaches how to hand over material, judge what comes back and keep control of your work, and none of that is specific to one assistant. Claude is named where a screen or a feature is being described, because a course has to show something concrete.
The days are the course, in order, with exercises. The 55 confusion pairs are standalone — two things people treat as one, and the wrong decision that follows, readable in any order. The 2,995 concepts are a reference you look things up in. The 270 rules are what goes wrong and what to do instead, sorted by when you meet it.
No certificate, and nothing is scored. Each day carries questions, and the wider view deck at the end of a day is offered rather than tested — take three cards or one and stop. The thing you leave with is the work you did on your own material, which is worth more in a conversation than a badge.
Do the first hour: what an assistant actually is, why answers come back generic, and finding a file you had given up on. Those three pages change how you ask more than anything else here. The rest of the course is what makes it a habit rather than a good afternoon.