# Thoughts on ankification

Anki is an incredibly powerful tool for augmenting human memory. It is easy to make cards, but deceptively difficult to make cards well. The art of “ankification” involves transforming newly learned concepts into a card structure which is both:

1. Effective — the cards trigger actual recall in some situation outside of SRS practice, such as finding the right word when speaking a foreign language, or recognizing a link between two concepts.
2. Efficient — reviews are challenging but not impossible, and you rarely fail to recall the concept being tested.

I have been using Anki for more than three years and my process and methodology around “ankifying” concepts has continued to evolve. Here are some thoughts from along the way on what worked well and what didn’t.

## What worked well 💪🏼

### Mental model note

A note template with fields: concept, definition, visual, implication, ELI5, drawbacks. I use selective card generation to to create the following cards if the corresponding fields are non-empty:

1. Definition → Concept
2. Visual → Concept
3. Concept → Definition
4. Implication → Concept
5. Concept → ELI5
6. Concept → Implication
7. Concept → Drawbacks

### Vocabulary note

Includes fields: target language, translation, definition, example, picture, audio, test typing, forms, audio (forms)

1. Comprehension: FL → EN
2. Comprehension: Audio
3. Production: EN → FL
4. Production: Definition
5. Production: Typing
6. Production: Forms

### Math proof note

This one only has as single card, with four fields:

2. Approach
3. Solution
4. Source

I have approach configured as a hint field so that I can reference it if I am struggling, without fully peeking at the answer. When this happens, I mark my recall performance as “Hard” rather than “Good”.

## What didn’t work well 😢

### Multi-concept math card

I had a note template which contained a concept and three prompts, each of which involved solving an isomorphic problem with different specific content. The prompt shown was determined by a piece of javascript which took the date modulo 3. This made it random enough, but ensured that if I got a card wrong, I would see the same prompt later that day.

Here is a somewhat contrived example:

Product rule (calculus): $(fg)^{\prime} = fg^{\prime} + gf^{\prime}$

Q1: What is the derivative of $x \sin x$? A1: $\sin x + x \cos x$

Q2: What is the derivative of $\sin x \cos x$? A2: $\cos (2x)$

Q3: What is the derivative of $\sqrt{x} \cos x$? A3: $-\sqrt{x} \sin x$

The motivation here was to force myself to recognize the underlying structure of the problem, rather than subconciously recognizing the specific values in the problem and remembering the solution.

Ultimately I abandoned this approach in favour of three individual proof cards, with the concept repeated in the “context” field of each. Is this less DRY? Perhaps. But these cards just didn’t perform well. I noticed that these multi-concept cards were disproportionately present in the list of “leech” cards which I had failed to recall repeatedly.

I suspect the reason is that the individual prompts were of slightly varying difficulty. If I struggle to recall prompt 1 today but ultimately succeed—therefore pressing the hard button on Anki—I would like to see that same prompt again soon. But on the next review, I only have a ⅓ chance of getting that prompt. This messes with the mechanics behind Anki’s ease algorithm.

### English vocabulary

I tried exporting my kindle dictionary lookup history and generating a set of flashcards using the same vocabulary template I used for foreign languages. This didn’t work very well, judged by the fact that a large proportion of these cards ended up with a “leech” tag after multiple failed attempts at recall.

I suspect the reason is that these English words were too nuanced for my existing vocabulary template, which includes fields for a picture, example sentence, and a short definition. Words such as stalwart, taciturn, finagle, or alacrity are difficult to recall precisely from a photo or even a definition. If I revisit this in the future, I suspect I’ll need to make use of cloze deletion rather than structured cards.