Revise before an exam

~ min read

30-second summary

Studying and revising are two different jobs. Studying = understanding. Revising = simulating the test until the answer comes out fluently.
Three techniques: simulated oral exam with a professor pressing you one question at a time, written test with correction on the step where you went wrong, mind maps and flashcards to fix structure and details.
The core prompt for the oral exam has two non-negotiable specifications: one question at a time, and an honest verdict on the answer. Without them, AI accepts vagueness.
The main risk of this method: AI that confirms you. It has to be forced to be strict, and the key points must be checked against the textbook.

You’ve studied the syllabus with the techniques from Study a new topic and composed the material with those from Notes from lectures and long texts. Two weeks left until the exam. You know the material, but the professor will ask you about it on your feet, under pressure, at their pace.

Studying and revising are two different jobs. Studying is understanding the first time. Revising is repeating what you already know until the answer comes out fluently under exam conditions. The textbook alone doesn’t generate those conditions: you open it, you read, you nod along. AI does: you can put it in the professor’s shoes and have it question you as many times as you want, for free, at three in the morning if needed.

The three techniques

1. Simulated oral exam

The central move of revision. The core prompt:

“Pretend to be my [subject] professor at the oral exam. Here’s the course syllabus: [paste the textbook index or the list of topics]. Ask me one question at a time, don’t move to the next one until I’ve answered. After my answer, give me an honest verdict (accurate / imprecise / missing), and if it’s imprecise tell me what you would have wanted to hear. Don’t move to the final grade until I ask you to.”

Three specifications are what matters. One question at a time: without this, AI fires four questions in a block, you answer sketchily, and you lose the pressure of the face-to-face. Honest verdict on every answer: by default AI wants to please, and without explicit instruction it accepts vagueness with a “good answer, let’s move on”. No final grade until you ask for it: without this, after three questions it closes with a B+ and you walk away happy; with the clause the questioning goes on until you decide, like a real oral exam.

The syllabus pasted into the prompt prevents AI from pulling topics from Wikipedia that your course doesn’t cover. If the textbook index is long, the list of chapters covered in class is enough. If the course has no single textbook (slides + handouts + a few articles, a common situation in social sciences and humanities), reconstruct the list of topics from lecture and slide titles. Fifteen to twenty concise entries are enough to define the perimeter.

Practical pace of the session. A good simulated oral session lasts forty minutes, five to eight questions. No more in the same round: after eight questions your attention drops and AI starts recycling similar questions. Two forty-minute sessions on different days work better than one two-hour marathon. You know you’re ready on a chapter when you answer three questions in a row without AI having to correct you on anything substantial; below that threshold, go back to the textbook before redoing the oral.

2. Written test with correction

For written exams, the pattern is the same but the loop changes: you produce a text, AI reads it and corrects you at the point where you went wrong, not in general.

“Give me a typical [subject] written exam prompt on chapter X. I’ll answer it, then you give me an honest verdict: where I wrote correct things, where I was imprecise, what I should have added. If I got a specific step wrong, show me the correct step, not just the final result.”

It works the same way for technical problems in math, physics, chemistry, engineering. The important difference is asking for the correction of the step, not just the result. Knowing that “the result is 12 and not 14” doesn’t help you avoid the same mistake again; knowing that you misapplied the chain rule at the third step does.

For multiple-choice written tests (fifty questions in two hours, A/B/C/D) the pattern shifts slightly: ask AI to give you ten or fifteen questions in the style of your exam, with four alternatives each and the answer marked. You answer before looking at the solution, and for each wrong one ask why the others were wrong, not just why the right one was right. Without the explanation of why not the distractors, multiple-choice becomes pattern memorization and you don’t grasp the structure of the question.

3. Mind maps and flashcards

Two tools complementary to questioned revision.

The mind map is a hierarchical diagram: main concept at the top, sub-concepts branching out. It’s useful before the oral exam to “see” the structure of the chapter at a glance. Ask: “give me a textual mind map of chapter X, main concept at the top, four or five branches, two or three sub-concepts per branch.” The output is an indented outline. If you prefer the graphical version, paste it into XMind, MindMeister, or Excalidraw.

Flashcards are the opposite of the map: no structure, many short questions on single concepts, to be repeated at growing intervals of time (the principle of spaced repetition: you review a card when you’re about to forget it, so each repetition costs less than the previous one).

Anki and Quizlet are the two standard tools. Anki is free, open source, and shows you the spaced repetition algorithm transparently; Quizlet is friendlier graphically, with a more limited free version. Both import flashcards from text files: the base format is Question,Answer (CSV) or Question<TAB>Answer (TSV), one card per line.

Ask AI: “give me twenty flashcards on chapter X in CSV format, one per line, comma as separator, no headers. If an answer contains commas, wrap it in double quotes. Focus on definitions, dates, names, and key formulas; avoid generic oral-exam questions.” You copy the output, save it as .csv, import it.

Real limit: AI tends to ask questions that are too easy (“Who was Kant?”) or too open-ended (“Explain the categorical imperative”). Check the first ten cards before importing everything; if they don’t work, send AI back with an example of a good question against a bad one.

Strategy for the professor

Two typical profiles. The techniques stay the same, the calibration changes.

The predictable professor. Every year asks about the same topics, with a recognizable type of question (precise definitions, applications to cases, links between authors). If you have a list of questions from past exams, even just the ones circulating in the course’s WhatsApp groups, paste it in: “this is the style of questions the professor asks. When you question me, simulate it.” It’s the same calibration as the third technique of Study a new topic, carried over into revision.

The unpredictable professor. Asks reasoning questions, not memory ones: knowing chapter 4 by heart won’t save you, they’ll ask you to link it to chapter 7. Here AI is useful for generating transversal questions: “give me five questions that connect chapter 4 with chapter 7, the kind a professor would ask to see if you’ve integrated the syllabus.” They won’t be perfect, but they force you to build bridges that you wouldn’t build reading linearly.

And if it’s the first exam with that professor and you have no past questions or hints about their style? Go with the unpredictable profile. It’s the most defensive strategy: you prepare for connecting questions, which require knowing the syllabus more deeply, and so they also cover the case where the professor turns out to be the predictable type.

A philosophy oral exam

Student of theoretical philosophy, exam in a week. The professor is predictable, insists on Heidegger and in particular on the concept of aletheia. First thing: set up the core prompt of Technique 1 with the course syllabus and the simulation starts. AI asks the first question, the student answers. The exchange that follows.

You

For Heidegger, aletheia is the Greek concept of truth. The word literally means "unconcealment", and for Heidegger truth is not the correspondence between a proposition and a thing, but the way being manifests itself. Truth understood as idea, in the Platonic sense, is a consequence of this originary manifestation.

Accurate on two points: the etymology of aletheia as "unconcealment" (more precisely a-letheia, privative of lethe, concealment) and the critique of truth as correspondence. Imprecise on the last part. You said that "truth understood as idea, in the Platonic sense, is a consequence of this originary manifestation". Here you trip on one of the points the professor always comes back to. Heidegger doesn't say that Platonic truth as idea derives from aletheia: he says it replaces it, and that this replacement is the first fall of Western metaphysics. For Heidegger, Plato is the moment when truth stops being disclosure (event, happening of being) and becomes adequation of a gaze to an ideal form. The originary aletheia is covered over by the idea, not prolonged in it. You should have said: "truth as disclosure is what Plato, according to Heidegger, loses by introducing the idea as the measure of the true". Not a small difference: in your formulation Plato is an heir to Heidegger, in the correct formulation he is the beginning of his polemical target. When you're ready I'll move on to the second question. Accurate on two points: the etymology of aletheia as "unconcealment" (more precisely a-letheia, privative of lethe, concealment) and the critique of truth as correspondence. Imprecise on the last part. You said that "truth understood as idea, in the Platonic sense, is a consequence of this originary manifestation". Here you trip on one of the points the professor always comes back to. Heidegger doesn't say that Platonic truth as idea derives from aletheia: he says it replaces it, and that this replacement is the first fall of Western metaphysics. For Heidegger, Plato is the moment when truth stops being disclosure (event, happening of being) and becomes adequation of a gaze to an ideal form. The originary aletheia is covered over by the idea, not prolonged in it. You should have said: "truth as disclosure is what Plato, according to Heidegger, loses by introducing the idea as the measure of the true". Not a small difference: in your formulation Plato is an heir to Heidegger, in the correct formulation he is the beginning of his polemical target. When you're ready I'll move on to the second question.

Three things to notice. The verdict isn’t a generic “good answer”: it’s split into accurate / imprecise, with the imprecise zone pointed out down to the millimeter. The correction doesn’t just say “no, it’s the opposite”: it tells you what you should have said, with the alternative formulation. And AI stops: it doesn’t move to the second question until the student asks. The simulation holds.

This is the difference between using AI as a revision tutor and as a serial question generator: the value isn’t the number of questions asked, it’s the detail of the correction on each one.

AI tends to agree with you. If you answer something vague but that sounds “in the ballpark”, the default response is “yes, you got the point”, followed by an addition that fills in the gaps you left. You walk out of the questioning convinced you knew, but you knew vaguely: the pieces you were missing, it put them in.

Three countermeasures. The first: in the initial prompt, explicitly ask for a strict verdict, not an encouraging one (“if I answer vaguely, tell me it’s vague, don’t fill in the gaps yourself”). The second, and most important: the key points where AI corrected you need to be re-checked against the textbook. An AI correction that sounds right to you can be a hallucination, especially on authors, quotes, dates. The same rule applies as in When to trust it and when not: cross-check with the source on the points you’ll bring to the exam.

The third, optional but useful: for the points you’re most unsure about, ask the same question to a second AI (Claude if you used ChatGPT, Gemini if you used Claude). If the two corrections agree, that’s a higher signal of reliability. If they diverge, you’ve found a candidate to verify first against the textbook. It’s not a guarantee (two AIs can hallucinate the same thing), but it raises the bar over a single source.

When to verify against the textbook, in real time or at the end of the session? For pointwise corrections (a date, a name, a quote) it’s better to jot them down on the fly during the session and verify them in a block afterwards. For big conceptual corrections (a thesis attributed to one author instead of another, a philosophical distinction) it’s better to stop right away and open the book: if you revise for an hour leaning on a wrong correction, you fix the error.

What NOT to do

Don’t revise on AI’s answers instead of on the textbook. An AI correction has to be verified against the book before you make it yours. AI confirms errors if you state them with confidence in the prompt: it tends to follow you.

Don’t import a deck of flashcards without checking the first ten. Wrong cards repeated with spaced repetition fix the error in place of the concept.

Don’t treat the AI’s “grade” as predictive. A simulated B+ means you had a good conversation for forty minutes with a system designed to accommodate you. The grade is there to close the session, not to make predictions.

Check what you’ve understood

What comes next

Revision is set. The next lesson changes register: the thesis, where what’s at stake isn’t a twenty-minute exam but a months-long project, with a supervisor in place of the exam committee.

About Privacy Open manual, written by a real person, in collaboration with AI.