Finals revision retrospective

When revising for my Finals, I found it quite useful (and interesting) to read former students’ writeups about how they studied for theirs, and now it’s my turn to return the favour. I’ve already published all my revision materials (and other academic work), plus a synthesis of what I took away from the degree, so I thought I’d round things off with a retrospective of how the revision actually went.

Prelude: some brief comments on choosing papers

Studying PPE at Oxford, you take all three subjects for the first year, and then have the option to drop one for second and third year. Oxford’s exam structure is a little unusual in that – for some humanities subjects, at least – the only thing that contributes to your overall grade is Finals, the set of exams you take at the end of third year. For PPE, that means the whole degree comes down to just 24 hours of work divided across eight closed-book exams. This system is relatively cheating-proof, unlike take-home assignments or dissertations which can be substantially AI-written or -assisted, and it also means that you benefit from synergies between same-subject papers, as you’re studying for them all at the same time rather than sequentially. Plus, you end up having a “Trinifree” term in your second year with no exams and plenty of time to enjoy the summer weather. But the obvious downside is that it’s an awful lot of concentrated pressure, and doesn’t leave you with much time for anything else besides revision in the few months leading up to Finals.

I took the philosophy-economics route for PPE, with a tilt towards econ: my papers were Ethics (compulsory), Aristotle’s Nicomachean Ethics, Philosophical Logic, Microeconomics, Quantitative Economics (introductory metrics), Econometrics, Game Theory, and Microeconomic Analysis. The papers I chose meant that I ended up doing only economics in third year, so it had been about a year since I’d written any essays when I started revising.

Scheduling & motivation

• Eight exams is a lot, and I definitely got bored of revision by the end – the exam period itself stretched over 3.5 weeks, and I had been revising for three months total. So it’s important to figure out what will be a sustainable pace.
  • For me, a target of 6h/day with Sundays off was the initial plan, and I mostly stuck with it (although once my exams started, I was flexible about when I’d take time off, based on when I had gaps in between).
    • You should use time-tracking software, because otherwise it’s very hard to get an accurate sense of how long you’re actually spending on revision. I recommend Clockify, but other people I know like Toggl or ActivityWatch.
  • In total, this added up to about 500 hours of revision, or roughly a quarter-year of full-time work.
• Having a revision routine with friends made things a lot more tolerable.
  • A friend of mine got together a few group revision classes for logic, and this was some very useful accountability to actually learn what I was missing (especially since the others in the group were very good at logic! One of them even going to do a masters in it…)
  • Merton set up a “study bunker” in one of the rooms in college, and I spent a lot of mornings there revising before lunch in Hall. Even if you’re not collaborating, coworking does make things a lot less boring.
    • (Sometimes the micro-interruptions get annoying, if you’re trying to really deeply work – and you do need some of that too.)
• I got Claude to write me a colour-coded HTML revision planner, which made life lots better.
  • One failure mode I was worried about was decision paralysis from not being sure about what to revise at any particular moment. To avoid that, I spent a bit of time (like, 2h or so) at the very start planning out roughly how long I wanted to spend on each subject and then each Sunday would fill out my schedule for the following week, in half-day blocks.
    • You might be interested in Tom Adamczewski’s blogpost on optimising your time allocation between subjects (and you could use the variance data from Finalcast if you wanted to take a similar approach to his).
    • I was a lot less rigorous: I just set time budgets based on how hard I was subjectively finding each subject, with a bias towards high-variance technical papers. QE and Micro had the least time; Logic had the most (3x as much as Micro), followed by Metrics and Micro Analysis.
  • You can see my planner template here, but you might as well just ask Claude to build you a personalised one exactly as you like.
• There’s so much sampling variation. Outcomes only depend stochastically on effort (and the relationship itself isn’t known!).
  • The easiest exams are often the least satisfying while you’re in them, because you’re mostly regurgitating material you’ve written before (or mechanically going through algebra, etc), but they do feel good afterwards.
  • On the other hand, when some of my exams went worse than I wanted / hoped, I took some time off afterwards to relax, rather than going back to revision.

High-level strategy

• Memorisation is useful for economics, because exams are often time-pressured and knowing what the answer is (or is likely to be) in advance obviously lets you reach it faster.
  • I think I overweighted some kinds of memorising early on, especially of complicated logic proofs and highly technical analysis definitions. But it’s hard to know for sure, because in principle they could’ve put that stuff into the exam.
• How do you figure out what to memorise? By looking at the past papers!
  • For each subject, I downloaded all the past papers from SOLO, transcribed them, and then used an LLM to identify common themes and frequency of different topics.
    • This was useful, and did lead to me changing some of my priorities.
    • But at some point you will just need to do the practice papers, and that’s probably still the best way to get a sense of what they’re actually like.
• I didn’t do any fully timed papers, and I think this is overrated. Practising questions timed is important, though.
  • For economics, I found it was more efficient to go through questions individually (under pro rata time constraints), and then review each one against the solutions immediately. That way it was fresher in my mind when I saw the solution.
  • For philosophy, doing timed essays had an ugh-field around it initially, but was actually quite fun once I got into it. However, doing too many timed essays is inefficient, because in the exam you will mostly be either (a) responding to the very specific question, or (b) reprising talking points you know.
    • So, I think the optimal strategy is to write e.g. four or five essays to get a feel for how much you can get onto the page in an hour (and polish your tone), but spend most of your time planning essays from past papers.
    • Before doing any of that, I spent a while writing up notes on each topic I planned to revise, as well as clearing up confusions I had (and clarifying what my own views were).
• Except for disagreements noted above, I defer to Hilary Greaves’s advice, plus this post.
  • One further point I’d emphasise is that, especially for normative ethics, having thought-experiments ready to deploy (or at least, practising the skill of formulating them) is quite useful. They can illustrate your point better than lengthy argument, and examiners seem to appreciate them.

More specifics on how I revised

The short version of this section: use flashcards (mine are here), practise past papers, read the exam reports, and think creatively about how LLMs can make your revision easier or more effective.

Anki flashcards

For logic and economics papers, doing flashcards in Anki was a substantial part of my revision – about 20% of their total time (plus all the time spent previously making the cards, about 30h per paper). I think this was probably about the right fraction: it meant that I knew basically all of the relevant course content fluently, but any more would’ve got in the way of me completing enough past papers. There’s a danger when making flashcards that you start to value having a beautiful deck as an end in itself, so beware of that!

Compared to when I was doing my Prelims revision, a lot of the process can now be automated, especially for writing technical flashcards (where a large amount of time is spent on tedious MathJax typesetting).

Here’s the workflow I followed for each econ paper:

1. Download all past papers and examiners reports for previous years.
2. Get Gemini to transcribe them all, along with all problem sheets.
3. Copy the transcriptions into Claude, along with the rubric / list of lecture topics for the course, and ask it to analyse what topics come up in practice, plus the ways in which they’re usually tested.
   • You could use a prompt like this:

     I'm studying for my game theory FHS exam. Can you go through these exam papers and help me get a better sense of what from the course to actually revise? e.g., for the week we do on repetition there's a lot about folk theorem proofs, but I'm not sure if this is actually necessary for the exams.
     
     I'd rather save the exam papers for later when I take them timed, so for now I only have psets to practise with but they're generally trickier than exams I think.
     
     I was thinking that you could go through all the exam papers carefully and synthesise what the main types of questions are, then next I'll upload all my psets, and you can draw your conclusions.
     

4. Copy that summary into a new chat, along with the slides from the first week of the course. Ask Claude to review the material and suggest what to cover within the flashcards. If relevant, get Gemini to transcribe handwritten notes from the lectures/tutorials, and add these too.
   • Usually I will draft my own rough list of desired flashcards first, because it seems like a good first step to just skim over the slides and remind myself of what was covered. Plus I don’t want to be totally out of the loop of making my own revision flashcards.
   • Example prompt:

     Can you go through this set of slides and help me find the main important things to do flashcards on? Cross reference with the attached transcript analysing past exams. I think some important things to cover are the following, but I might've missed stuff:
     - ...
     Also attached are some notes I made to myself, trying to understand this all better (because the slides felt quite confusing).
     
     [Other information that is useful for Claude to know -- for example
     (i) I will spend a while practising questions for this topic so it's not necessary or useful to memorise lots of derivations -- but probably what _is_ useful is some rough notes on general procedures, what to expect about the conclusions, the key equations to recall.
     
     or (ii) I found this topic tricky and will probably rely on mechanically following steps to the answer more than usual, vs deriving things from first principles.
     ]
     
     Don't start making the flashcards yet, just suggest what they should be & sketch out the content in a pedagogically-informed way.
     

5. Review the proposed topics, use judgement to decide whether any need to be reframed / dropped / added in.
6. Move the chat into my “Economics flashcards” Project, where there are further instructions for Claude about how to format the output, as well as more high-level guidance about my flashcard design principles. Once the list looks good, ask Claude to draft all the flashcards.
7. Copy each card over into Anki, and make any necessary edits to it. Or, if there are substantial changes needed, ask Claude to do that first.
   • The most common change I need to make is splitting up a card containing several concepts into separate ones.
   • Sometimes I won’t fully recall the intuition behind why a result is as it is; often I’ll then try to properly understand it in the moment but not bother to add this into the flashcard, since it pays to be a bit of a ruthless prioritiser for what to actually memorise for exams vs know you understood once (for the satisfaction).

Ideally you’d just give everything (raw exam PDFs, examiners reports, slides, and exports of handwritten lecture & tutorial notes) to Claude all in one go, but this really doesn’t work with current context limits. (I was doing this in web chat, and possibly Opus 1M would handle it in Claude Code / via API, but setting up the pipeline seemed like a possibly-fun-but-time-wasting distraction.)

Exam reports and practice questions

• Every year, the course convenor or another assessor of each philosophy paper writes an examiners’ report about how candidates did. Often these are extremely detailed – for all three of my phil papers, they had question-by-question comments about the typical responses and what they lacked. You should read them!
  • The exam reports are useful in at least two ways. First, they give you a general sense of what the examiners are looking for – their pet peeves, things that make scripts stand out to them, and so on. Second, they’re useful as “indicative content” about the specific moves you might make when answering individual exam questions.
  • Since the exam reports are somewhat spoilers for the actual questions, it’s better not to look at them until you’ve had a proper attempt at writing or at least planning the essay in question – although there’s no problem with reading the general remarks first.
• Get feedback on your essays, from both your tutors and LLMs.
  • Feedback from (good) tutors is most useful, since they have a far better sense of how Oxford Finals essays are marked than any LLM does.
  • However, they take time to respond, and also are unlikely to be willing to give you comments on tens of essays (or to engage in extended back-and-forth discussion about each one). So, it’s worth supplementing their feedback with Claude’s.
    • I set up a Project for each paper, and uploaded my master notes into it, as well as Hilary Greaves’s essay advice and the concatenation of all historic exam reports for that paper.
    • Then, I’d simply paste in my essay or plan(s), and ask Claude for feedback.
    • Here’s the prompt I used:

      I'm revising for Ethics finals following Hilary Greaves's advice. I'm planning answers to these past paper questions using my notes.
      
      Be aware that the examiners' reports are neither exhaustive nor stipulative; they are **indicative** of what a good response might include. (However, they may also include information about pitfalls to avoid, or -- when flagged explicitly -- _crucial_ ideas to mention in a response.)
      
      When providing feedback on the plans, I'd like you to help me consider:
      
      - Does my response fully answer the question as stated?
      - Are there points in the exam report I have omitted / fallen astray on?
      - Are there comments in my notes which I could have included?
      - Are there other relevant ideas or arguments to flesh out the answer, if I noted it being light?
      
      Bear in mind that when asking for feedback on essays or commentaries, these were done under timed conditions without any slack. So it's less useful to suggest new additions if you don't also suggest what you'd cut to make space/time. [But it is nonetheless still useful overall to suggest things I could've included, e.g. if another thought I had in mind didn't come to me, etc!]
      
      Be realistic when talking about what to cut. The words are not the constraint so much as thinking time and figuring out how to explain stuff. Once you have decided on presenting a thought experiment, for example, it's very easy to write it out. On the other hand, adding a new point to the dialectic is much more time-consuming -- and riskier, since if you don't manage to complete it, the essay doesn't work.
      
      Please review the following essay plans for me.
      

  • Overall, I think current models (e.g. Claude Opus 4.6 or 4.8) have substantially worse taste than I do about what’s philosophically interesting or important to include in an essay, even with full access to all the above materials. Despite that, they’re still useful for critiquing my work, since they’re often able to provide suggestions of arguments I’ve missed, or ideas in my notes that could’ve been incorporated, and so on.
    • Something to beware of: I would once or twice each day catch Claude making some kind of philosophical error, and presumably I was only noticing a fraction of these. They were fairly subtle points, for the most part, but the kinds of things that an examiner would probably pick up on.
• For technical questions, if there are no solutions available, you can get a mark scheme by just giving the problem to an LLM and asking for the answer.
  • I found that Gemini Pro was generally faster than Claude Opus and had fewer errors, especially for game theory and logic.

In the exam

• Don’t let physical discomfort be a distraction! Although you need to wear sub fusc into the exam room, there’s no reason to keep it all on once you’re inside. I would always remove my gown and bow tie, and undo my collar (plus roll up my sleeves) before starting.
• Take a cheap Casio watch with you – often you can’t see the clock(s) clearly, and you don’t want to misread them under time pressure, especially in papers where discipline is important, like Micro.
• Write something for every part of every question! You’ll get some marks, even if you’re just saying how you would try to answer it if you had time (or why your answer looks wrong, etc).
  • This is maybe the biggest difference from technical exams in school – there, everything is point-marked against a prescriptive mark scheme, so either what you’ve written gets those discrete points or it doesn’t.