This is a guide for after installment.

There used to be an excellent page walking you through how to set up Anki but the website is down. Here’s what I got out of it though:

• You should only have one deck because having multiple decks makes superficially easier to remember cards when you know which broad category your answer should be in. But of course if you are only using Anki for class then you can separate your decks.
• There are four types of cards.
  • Basic: Questions (e.g. what is the radius of the Earth?)
  • Basic and reversed: vocab, concepts (e.g. heat equation, anachronism)
    • This generates two cards. One shows you “anachronism” and asks you to recall the definition. The other shows you the definition and asks you to recall the word. This can help avoid tip of the tongue phenomenon.
  • Cloze: Sentences or paragraphs with multiple fill in the blank answers (e.g. Africa is splitting into two continents at a rate of ____. The two plates are called the ___ and ____.) Often useful when you copy paste from wikipedia.
  • Image occlusion: diagrams you want to learn how to label

Once that’s done and you’ve started using Anki, you can finetune your setup using the approach in Refold’s Basic Anki Setup Guide. These were my main takeaways from the guide:

• Only use the “Again” and “Good” buttons.
  • Not in this guide, but this post suggests that you click “Again” immediately if you don’t know the card rather than effortfully trying to recall.
• Use FSRS
  • “FSRS optimizer uses machine learning to learn your memory patterns and find parameters that best fit your review history”
• The most important statistic (see Stats) for your deck is the retention rate of your mature cards, which you ideally want to be 80% to 90%. If it is lower, then you are forgetting things you mastered earlier and if it is higher, then you are wasting time trying to perfect your cards.

Update 12/08/24: I used the above techniques for a while, and they were great but not sufficient to resolve many issues. Some things I still experienced included

• Having too much detail in my cards, which encouraged me to memorize the way the card looked or some cue that didn’t actually engage with the material
• Memorizing the teacher’s password (see LW post). I had memorized many definitions and algorithms thanks to Anki, but I was missing important metadata about how much credence to have in xyz theory, how an equation was derived, why something works in a certain way, etc. Example: I added holobiont theory to my deck thinking it was widely accepted, but realized in a class discussion that it was in fact heavily contested. This was very confusing and exposed something unhelpful about my Anki practice. In retrospect, this seems to be a fault of both not diving deep enough and the source material I was drawing from (e.g. academic papers, which often have their own agenda). I am honestly not sure if Anki is great at dealing with this type of knowledge. Can you even embed information like credences and takes into cards? More broadly, how deep should you go before you can trust that you’ve exhausted controversy and are now dealing with facts?
• Feelings of completionist duty, which got in the way of being excited about processing source material

Andy Matuschak’s How to write good prompts: using spaced repetition to create understanding seems helpful for addressing these issues (I have not implemented yet, these notes are being written as I’m reading the post). One nice point was that in addition to being a memory retrieval tool, Anki effectively tries to extend and control the effects of Baader-Meinhof phenomenon.

• If you are making very detailed cards, you won’t be fully stimulating memory retrieval. You want atomic cards that produce consistent answers every time
  • For cloze cards, focus on only one detail at a time
  • I often make many cards about one topic in a burst. Provide context at the beginning of each card instead of burdening your future self with recognizing the context. E.g. in a burst of cards about SCNT, start the card with “when performing SCNT, …”
• Add more cards to get the metadata. Some questions might be:
  • Ask for examples of the concept
  • How does this concept differ from another similar concept?
  • What is the status of this approach or theory in the literature? Why do some people not believe it? What is getting in the way of not resolving the controversy?
  • When should I use this framework as opposed to some other framework?
• There are multiple types of LISTS.
  • Open lists aren’t exhaustive but share same general characteristic. If you are tempted to write a card with an open list, you might want to try extracting the shared characteristic to approach some generality: when I observe this _____, what should I think of? (this is asking the opposite of under what conditions should i apply this framework)
  • Closed lists are exhaustive and can be either ordered or unordered.
    • If you have an unordered closed list, use one Cloze at a time so you can figure out what’s missing. Then see if any of the steps can be canceled out because they are implied by others (e.g. if you are boiling water, presumably you could figure out that the first step was pouring water in). Write the cards in the format: when doing [context], if I am at step such and such, what should I do next?
    • If you have an ordered closed list (e.g. procedure or algorithm), write out list as minimally as possible.
• CREATIVE vs RETRIEVAL prompt: Creative prompts ask you to come up with a new answer you haven’t used before for an open list. This seems like it would be a chore. Not going to use this
• Most importantly, it’s hard to tell how a card might fail when you make it. Iterate and edit!