Every study guide gives you the same useless advice: "start early, review consistently, don't cram." Great. But sometimes it's two days out, the exam covers fourteen weeks of material, and "review everything" is not a plan — it's a way to run out of time having studied the wrong things.
Crunch mode needs a different premise. You're not trying to learn the whole course. You're trying to maximize expected points per hour. And the single best predictor of what an exam tests is what previous versions of that exam tested. Professors reuse structure. They have topics they care about, question formats they favor, and a point budget that reveals what they think matters. That signal is sitting right there in past exams — you just have to extract it.
This is a task AI is genuinely good at: reading a stack of documents, tallying patterns, and turning them into a schedule. Here's the workflow.
The Premise: Exams Are More Predictable Than the Syllabus
A syllabus is a promise about coverage. An exam is a revealed preference about priorities. Those are not the same document.
A course might spend three weeks on some foundational topic and one lecture on a flashy application — and then put the flashy application on the exam every single year because it makes for a clean question. If you study by syllabus weight, you over-invest in the three-week topic. If you study by exam weight, you catch the pattern.
The goal of this workflow is to recover that exam weight: which topics recur, how many points they're worth, what format they show up in, and where the easy points hide.
What You Feed It (and What You Shouldn't)
Before anything else: use materials you're actually entitled to. Professor-provided practice exams, official past papers a department posts publicly, sample questions in the textbook, exams your instructor handed back to the class. Plenty of courses publish several years of finals precisely so students can practice on them. That's the corpus.
Don't use anything you're not supposed to have — leaked exams, a current exam, or another student's private files. The whole point is that legitimate past exams are already a strong signal. You don't need to cheat to benefit from this; you need to read what's in front of you more systematically than you have time to do by hand.
Gather what you've got — PDFs, photos of printed exams, the practice set — and drop them in a folder. Three or four past exams is enough to see patterns. Even one, analyzed well, beats studying blind.
Phase 1: Extract and Normalize
The first job is turning a pile of documents into structured data. I hand the AI the exams with a brief like:
"Here are several past finals for the same course. For each exam, list every question: its number, the topic it tests, the point value, and the format (multiple choice, short answer, proof, calculation, essay). Give it back to me as a table."
This sounds mechanical, and it is — but doing it by hand across four exams is exactly the kind of tedious pass you'll skip when you're panicking. The AI reads all of them in one shot and produces a clean per-question inventory. Photos work too; modern models read a picture of an exam page fine, though it's worth spot-checking the point values it transcribes.
The output of this phase isn't a study plan yet. It's just the raw ledger: every question that has ever appeared, tagged and priced.
Phase 2: Find the Point Distribution
Now the analysis. With the inventory built, ask it to aggregate across all the exams:
"Across all these exams, group the questions by topic. For each topic, total the points it's worth on average per exam and its share of the total. Show me which topics appear on every exam versus occasionally. Flag any topic that's consistently high-value."
This is where the shape of the exam appears. A few things reliably fall out:
The load-bearing topics. Usually a handful of topics carry the majority of the points, and they show up on every past exam. These are non-negotiable. If a topic is worth 25% of the points and has appeared four years running, it will almost certainly be worth 25% of the points this year.
The reliable freebies. There's often a category of easy, formulaic points — definitions, a standard derivation, one particular calculation done the same way every time. Low effort to lock in, and they always appear. In crunch mode these have an outrageous points-per-hour ratio.
The long tail. A scatter of topics that showed up once and never again. Individually low-probability. Collectively they're where you stop spending time, not where you start.
Ask the AI to be explicit about confidence: which patterns are strong (appeared every time) versus weak (appeared once). You want it to distinguish "this is on every exam" from "this happened once in 2022." The difference determines where your hours go.
I also find it useful to ask it to note format trends. A topic that's always a multiple-choice question needs different prep than one that's always a from-scratch proof. Recognizing a concept and being able to produce it cold are different skills, and the format tells you which one you need.
Phase 3: Turn Points Into a Schedule
Analysis is worthless if it doesn't become a plan. The last phase converts the distribution into a time-boxed schedule against the hours you actually have:
"I have 12 hours before the exam. Using the point distribution above, build me an hour-by-hour study plan that maximizes expected points. Prioritize high-value recurring topics and quick-win freebies. Tell me explicitly what to skip."
The "how many hours" input is what makes this crunch-mode planning rather than generic advice. Twelve hours and forty hours produce very different plans — and the AI will make different cuts for each. Give it the real number, not the number you wish you had.
A good plan that comes back has a few properties:
- It front-loads the highest-value recurring topics. The stuff worth the most points that appears every year gets your freshest hours.
- It time-boxes each block. "Two hours on topic X" not "study topic X." Boxes stop you from disappearing into your favorite topic and starving the rest.
- It explicitly names what to skip. This is the part people won't do for themselves. Deciding not to study something feels like giving up. Coming from an analysis of the point distribution, it feels like strategy — because it is.
- It schedules practice on the actual format. If proofs are 40% of the exam, the plan has you writing proofs, not re-reading them.
The best final step is to have the AI generate a short practice set in the exam's own format — questions on the high-probability topics, in the style the past exams used — so your last blocks are active recall under something like test conditions, not passive review.
What Works About This
It replaces dread with triage. The paralyzing part of crunch mode is not knowing where to start. Fourteen weeks of material is undifferentiated fog. A point-weighted priority list cuts through it: here's what matters, here's the order, here's what to ignore.
It's honest about probability. Studying is an allocation problem under uncertainty, and this workflow treats it like one. You're not guaranteed the high-frequency topics show up — you're playing the odds, deliberately, with the odds made explicit.
The AI does the part you'd skip. Nobody hand-tabulates four exams into a point-distribution table at 11pm. But that tabulation is exactly what turns a vague "I should probably know this" into "this is 22% of the points." The tedium is the value, and it's the tedium the model is happy to absorb.
It generalizes past exams. Certification tests, licensing exams, standardized tests, coding interviews — anything with a published bank of past questions and a stable structure responds to the same treatment. Wherever the format is predictable, past instances predict future ones.
What It Won't Do
It won't teach you the material. This is a prioritization tool, not a tutor. It tells you what to study and in what order; you still have to do the studying. If a topic is 30% of the points and you don't understand it, the plan just means you'll spend your time in the right place — not that the points are free.
It can be wrong about this year. Professors change exams. New topics get added, formats shift, a favorite question retires. Past exams are a strong prior, not a guarantee. Treat "appeared every year" as high-probability, not certain, and keep a little time for the topics you know were emphasized in class but haven't shown up on a past exam yet.
Garbage in, garbage out. If your corpus is one unrepresentative practice exam, the distribution it produces is noise dressed up as signal. More exams, and exams that actually resemble the real thing, make the prediction better. Be skeptical of a confident-looking table built from thin data.
It doesn't excuse leaving it to the last minute. The honest caveat: the best version of this is run a week out, when the plan still leaves room to actually learn the priorities it surfaces. Two hours before the exam, even a perfect priority list can only do so much.
The Takeaway
Crunch mode isn't about working harder in less time — it's about being ruthless with where the time goes. The mistake is spreading a fixed number of hours evenly across everything the course covered. The fix is spending them where the points actually are.
Past exams already encode that information. What the AI adds is the willingness to read all of them carefully, tally the points without cutting corners, and turn the result into a schedule you can follow when you're too stressed to plan clearly yourself. The model handles the bookkeeping. You handle the studying — just aimed at the right targets.
You still have to know the material. But at least you'll be out of time on the right topics.