I enrolled at university at 41, while working, and sat every exam of the three years in two. No credits recognised for experience, no shortened track: a rhythm, a method, and an AI used as a tutor rather than a shortcut. The rhythm is one case, not an average, and it depended on twenty years of business in the business subjects. The method, on the other hand, can be copied: this page shows how it works, and the Quaderno 3in2 (twelve pages, in Italian, I send it to you) also has the five prompts word for word.
1. People study by rereading, and it does not work
Rereading and highlighting give a feeling of mastery that vanishes at the exam. Time has to move from reading to being questioned, and AI is the cheapest machine ever built for being questioned.
What little I studied at school I studied by reading, underlining, rereading. At 41 I found out that the method does not hold, and it is not my impression. Roediger and Karpicke showed in 2006 that people who retrieve information through tests remember it better over time than people who reread it, even though rereading feels more effective while you study. Dunlosky and colleagues reviewed ten study techniques in 2013: retrieval practice and spaced repetition came out as high utility, rereading and highlighting as low utility.
Translated: you learn by retrieving information, not by looking at it again. And this is where AI comes in. Not because it knows the subject better than the book (it often knows it worse, chapter 7), but because it questions you twenty-four hours a day, without tiring, on your own materials.
A mistake I made. I went to the marketing exam without studying, because in twenty years I had done and paid for a great deal of marketing. I came out with 18 out of 30. I retook it: 30. The subject in which you feel like a king is the one you estimate worst.
2. The six-move loop
Every chapter goes through the same loop, and you do not move on until the loop is closed. You go back to the first move only for mistakes of knowledge: all the others are cured with drills and mocks, not by rereading.
| # | Move | What happens |
|---|---|---|
| 01 | Explain | The AI explains one block of content, anchored to the materials you uploaded, with the source cited |
| 02 | Checkpoint | Right after the block, three to five quick questions. Inside the lesson, not at the end |
| 03 | Drill | A battery of questions with realistic traps on that chapter. Score and mistakes |
| 04 | Mock | A timed exam simulation, real format, several chapters together. Score and time per question |
| 05 | Diagnosis | Every mistake classified with a code (chapter 6). The log is updated |
| 06 | Re-drill | New questions only on the types of mistake that came up, until the mistake is closed |
If you work. The time was in two places: fixed blocks (two hours in the morning on one subject, one hour in the evening on another) and gaps (the hours on planes with the laptop, the hours in the car revising out loud with the AI). Not the heroic Sunday hour: a little every day.
3. How to build a mock that is worth something
The simulation is the move worth the most, and also the one the AI gets wrong most readily if you give it no rules. These are mine.
- Same format as the real exam: type of questions, number, scoring, time. You describe the format, the AI replicates it.
- Real distractors: every wrong option must be plausible on a quick read and use the course’s vocabulary in the wrong context.
- Correct answers spread across all the letters. AIs tend to pile them on one: check it.
- Options of similar length: the right answer must never be the longest.
- Difficulty above the real exam. If the mock is easier than the exam, it is useless.
- Timed, with the time per question recorded. Slow questions say more than wrong ones.
- A rationale for every answer, shown only after you hand in.
- Calculations verified before trusting the key: by hand, with a spreadsheet or by having the AI run code.
The mock is done in real conditions: phone in another room, timer, no breaks. Twenty-five real minutes are worth more than two hours in bits, and they are also easier to find.
4. The tutor’s instructions
This is the prompt that matters most: eight rules pasted into the project instructions, once per exam. The Quaderno has it word for word, with the parts to fill in.
The role first: you are my tutor for this exam, your job is not to study for me but to question me and correct me. Then the rules. Every explanation starts from the uploaded materials, with file and page reference. “From your materials” and “general knowledge” are always kept apart. If something is not in the materials, it says so and does not reconstruct it. Multiple-choice questions have the correct answers spread across all the letters, options of similar length and distractors that use the course’s vocabulary in the wrong context. Calculations are checked twice. When I get something wrong, no answer: a question that leads me to find the mistake. At the end, the exam format, my starting level and the hours I have each week.
What to upload and what not: your slides, your handouts, your notes, yes. Documents with other people’s personal data and materials your course rules forbid sharing with external services, no. And check the data-use settings of your account before uploading anything.
5. Lesson, drill, mock
Three prompts for the first four moves. The traps in the drill are requested explicitly, because left alone AIs ask questions that are too kind.
The lesson with checkpoints is move 1 and move 2 together: it explains a chapter using only the uploaded materials, in short blocks, and after each block it stops and asks three quick questions; it does not move to the next block until I have answered, and at the end it says which concepts I do not master.
The drill is move 3: multiple-choice questions harder than the exam, with the traps requested by name (at least two “which one is NOT”, one “find the false statement”, one changed word, “none of the above” only where it really is right), one question at a time, “correct” or “wrong” and nothing else, rationales and summary only at the end.
The mock is move 4: all the questions together, in the real format, with at least two that mix chapters, no solutions, the timer starts when you see them. On handing in: total score and score per section, a rationale for every question and, for every mistake, a proposed code. Before showing the questions, the AI checks that the correct answers are spread across the letters.
A mistake I made. In one mock the correct answers were almost all on the same letter. A mock like that measures nothing, and after three questions it teaches you the letter instead of the subject. Rule 4 of the instructions is there for this.
6. Not every mistake is cured by studying
On a multiple-choice and calculation exam, at an advanced stage of preparation, most of my mistakes were not about knowledge: they were execution under pressure and time. Studying more does not help someone who fails like that. Simulating more does, and to know it every mistake has to be classified.
| Code | Type | The signal | The cure |
|---|---|---|---|
| K | Knowledge | I did not know the concept | Back to move 1 on that point, then drill |
| E | Execution | I knew it, but got the calculation, sign or formula wrong | Timed numerical drills on the same pattern |
| L | Reading | I misread the question (NOT, except, units) | Underline negations, reread before answering |
| T | Time | I knew it, but did not finish or guessed | Mocks with reduced time, a skipping strategy |
Two more codes: D if you fell into a distractor built on purpose, A if the mistake was the AI’s (wrong key or explanation). The log is a table. Mine had these columns, and they were enough.
| Date | Exam | Topic | My answer | Correct | Code | Cause in one line | Closed |
|---|---|---|---|---|---|---|---|
| 12/03 | Accounting | Inventory | B | D | L | Missed the “NOT” | yes |
| 12/03 | Accounting | Depreciation | 2,400 | 4,800 | E | Halved the charge for no reason | no |
Diagnosis and re-drill, moves 5 and 6, share one prompt, to be used after every mock with the mistakes pasted one by one: for each one the code and the reason, the cause in one line, a concrete countermeasure; then the dominant type of mistake and a re-drill of ten questions on that type only. The five prompts, word for word, are in the Quaderno.
7. Three rules for not being fooled by the AI
I was about to memorise a wrong formula, beautifully explained. Being able to catch the mistakes of a tutor who speaks with confidence is an exam skill: it forces you to know the source better than it does.
The invested capital formula, in a version incomplete compared with the course’s: I noticed by comparing it with the slide. It was not the only case: an answer key that marked the wrong answer as right, a “plausible” definition that differed from the slides, a mock with the correct answers almost all on the same letter. From there, three rules.
- Every exam-grade statement has a source. File, page or slide. If the AI does not cite it, you ask. If it is not in the materials, the AI must say so and you do not memorise it.
- Definitions are checked against the source before you learn them. Always. A “plausible” definition is the most elegant way to get an open question wrong.
- Every now and then the AI audits itself. You ask it to reread everything it has explained to you, compare it with the materials and list where it said something that is not there, where it used a term different from the course’s, where it simplified a formula.
It applies at work too. If the AI prepares an analysis and nobody goes back to the source, the incomplete formula ends up in a board deck with the same confident tone it reached me with. The rules are the same.
How it was built
The PDF comes from a Markdown file of twelve sections, one per page, laid out as A4 HTML on the tokens of my design system (Figtree for text, Spline Sans Mono for prompts and numbers, the site’s two backgrounds) and printed with Chromium: the file embeds only those two fonts. It is the same engine I use for my technical notes, so the document you download and the page you read have the same face. The prose is a little over two thousand words, prompts aside; the prompts are the ones I use, rewritten without references to my own course and tested on a demo project with an accounting handout invented for the purpose, four chapters and eight pages, to check that any AI cites the page and stops at the checkpoints. No university material goes into the document, by choice and out of respect for the people who wrote those materials.
Where to start, today
You do not need to read anything else. You need to do three things, in this order, by tonight.
- Open a project in your AI, name it after the exam, paste the tutor’s instructions (the first prompt page of the Quaderno) and upload the materials of the first chapter.
- Run a lesson with checkpoints on that chapter. Answer in writing, without looking at the slides.
- Run a drill of ten questions. Log the mistakes with their code. You have just closed the first loop.
Who it is not for
It is not for anyone looking for a way to pass without studying, for anyone who wants to use AI during the exam, for anyone looking for a course on a specific exam at a specific university. AI does not study for you, and during exams it is not used.
The Quaderno I send you myself: leave your name and email in the form below and it arrives in a few minutes. I use the address only for that.
If you have already run the first loop, write and tell me how it went. I read everything, and if I can help I answer. If I cannot, I tell you.
