a) I love Spaced Repetition systems mainly bc it feels empowering that you actually have a choice on what you will remember b) I cannot wait until there will be a fully AI driven one. I just want to chat via voice about concepts, and the agent attacks the question from different perspectives as opposed to asking in a statically formatted question-answer format c) I want that bc it will be nicer to just talk and bc the biggest problem with the system rn is that you memorize the static question–answer pair, association that is often tied to the context of your deck, and it’s often hard to use the knowledge in the real world. e.g.: I remember having in my deck a card that uses these words, asks in this specific way about something and I remember correctly that the answer is X. I think the only way I know out of this dynamic is to do more than one card per concept and to do reverse cards, but this is a lot of maintenance and in practice I rarely do it. If you have any pointers on how to deal with this please let me know!
I think the only way I know out of this dynamic is to do more than one card per concept and to do reverse cards, but this is a lot of maintenance and in practice I rarely do it
There is another way, which is to try prototyping this out! The implementation seems pretty simple, but actually finetuning (disambiguated) it to be useful would take time I’m sure. But maybe it’s an upfront cost..
I asked Opus 4.8 to see what are solutions to this in the two texts i liked about Anki https://www.lesswrong.com/posts/7Q7DPSk4iGFJd8DRk/an-opinionated-guide-to-using-anki-correctly and https://augmentingcognition.com/ltm.html The two additional solutions are: – make cards very atomic (compare a more complex one-card-anki to four-cards-anki with the same knowledge, but a lot more atomic—atomic wins bc this type of knowledge is quicker to retrieve and more adaptable to the real life context, e.g.: in real life context you may need only 2 of 4 atomic bits, and you need to retrieve cards really quickly, and perhaps reassemble with different knowledge bits) – to formulate cards from an angle of the real-world trigger you will most likely have.
----- Claude Opus 4.8 (full-answer) -----
Both texts address your exact problem (recall getting welded to the deck’s phrasing and context). Here is everything they offer, grouped by solution.
From “Guide on using Anki correctly”
Name the card for its real-world prompt, not its content. The core fix for “I only recall it when asked in that specific way.” Ask what real-life trigger should surface this knowledge, then build the card around that trigger. Their example: don’t name a card “length of whale intestine,” name it “animal fun facts,” because the real moment of retrieval is wanting an animal fun fact, not being quizzed on whale anatomy. If the natural real-life prompt genuinely needs a longer card, make the longer card.
Put the meaning in the answer, not just the bare token. The “24%” example: instead of “In 1950 Germany, what % of workforce was in agriculture? / 24%,” write “historical size of agriculture sector / 1950 Germany: 24% of workforce.” You are more likely to reach for “what do I know about the history of agriculture” than the narrow framing. Memorizing the isolated number without what it means is useless for real use.
Anchor to semantics, not visual shape. Their diagnosis of why cloze deletions fail is your exact problem: excessive context makes you learn “visual shape of the card leads to answer” instead of “semantic meaning leads to answer.” The problem is context volume, so strip context down until the recall path runs through meaning.
Redundancy as a deliberate feature. This generalizes your reverse-card instinct. Multiple cards with slightly varied prompts (small “epsilon” changes to wording and angle) train you to recognize the whole solution space rather than one frozen pairing. Cards should reinforce the pattern “circumstances lead to solution,” which is what makes knowledge fire in the real world.
Brevity forces transfer. Max ~9 words for most cards, at the absolute most 3 bullets / 18 words. If you cannot compress it, you do not understand it well enough to break it down. Short cards recall faster and resist context-anchoring.
From “Augmenting long-term memory” (Michael Nielsen)
Atomic questions you reassemble in unexpected ways. Break each idea into its smallest pieces. His soft-link example splits one routinely-missed card (“create a soft link”) into “what’s the basic command” and “what order do the arguments go.” The payoff he names directly: you later “assemble the atomic questions in an unexpected way.” Atomicity is what lets knowledge recombine in novel real situations instead of firing only as one block.
Never orphan questions, and add several per concept. This is the stronger version of your “more than one card per concept” instinct. He makes it a rule to never add just one question, always at least two, preferably three or more, so the fact becomes “the nucleus of a bit of useful knowledge.” Lonely orphan cards get missed constantly and are a waste. A tightly interconnected web of facts retains and transfers far better than isolated ones.
Multiple passes, 5 to 20 questions per source. Re-read hard material several times, adding questions after each pass. This layers the concept from different angles, which is a manual version of the “attack from different perspectives” agent you want.
No yes/no questions (from his work-in-progress notes). They let you pass on recognition without real recall.
I just created Dig Text, a new writing standard, which allows you to read the shortest version first, and dig deeper only where it interests you. I thought about dig about ten years ago, lol, and now, with the help of Claude Code, I spent two weeks thinking it through and coding it. It’s still in BETA, so I’m gathering feedback, but I would love if you try it and let me know what works and what doesn’t. https://digtext.github.io
I had a similar thought for math proofs, where you should be able to see an overall proof sketch first, then expand on any point to see the intuition/longer proof in a hierarchical way.
IMO there should be a variant of text expansion where the original text gets replaced by the longer version.
The “Plus” icon is an improvement over having a new paragraph / indentation, but still too distracting, especially the purple color.
Keyboard navigation is a bit clunky.
Pressing Enter should move focus somewhere sensible, rather than eating the focus. Probably to the ending parenthesis for expansion, and to the original + icon for collapsion.
Maybe just me, I expect ArrowRight on the Plus icon to expand the text too.
Let me use ctrl+z ctrl+y to undo / redo
Would be nice to have a “default expand n layers deep” button.
Please show me an example of ((text)) and :::dig::: tagsin the demo.
Double parentheses syntax
I like it
You are not allowing double parentheses to be escaped, try \((a))
Your parser is wrong, try ((a((b))c))
btw the selection of text in your demo is fucked (Firefox; Windows if that matters), it is selecting text one line lower than where my cursor is.
I am slightly annoyed that the Plus icon is not centered properly. It is a little bit too low.
The way your examples are structured with one sentence expansions seems bad. I don’t want to spend a few seconds clicking the button just to read one more sentence
It makes me want to have some sort of preview (full preview or how many levels / characters hidden)
Marginally related on the “dig deeper” part:
These pages uses side panels for Obsidian-style page linking. (via)
Thanks for your patience, I am back from my time off.
You listed a lot of bugs, visual problems that I want to fix. Thank you.
I agree that just text color for pluses may be a better choice here.
I agree that the best use is when digs don’t hold a small amount of text, but this is up to the writer
“Pressing Enter should move focus somewhere sensible, rather than eating the focus. Probably to the ending parenthesis for expansion, and to the original + icon for collapsion”. I think this is very interesting. I have been thinking how to approach this. I think there are tradeoffs here, mostly simplicity. There is a way, actually, to have preview text (see dig example and text in parenthesis). I will think if there is a sensible way to bake it in more universally into the dig format eg.: • {{preview text}} or similar.
What do you mean by preview text? Do you mean Here is some text that is shown ((and some text that is hidden))? I don’t consider that as preview text, I am imagining text showing up on hover (unsure about exact design). If you mean something else, I don’t understand.
Also I still have no idea how the :::dig syntax work.
hover don’t work on mobile so its not a viable solution IMO
how would do u feel about the solution from the image?
I didn’t fully thought :::dig syntax through but I imagined adding a script on top of the site and inside dig tags all bulleted lists would render as dig text. Am I missing something in this solution?
I like the idea, though I also already really like classical options such as footnotes, and don’t even mind footnotes on footnotes.
One UI feedback point: when expanding an item, it is currently shown exactly like the main text. I’d have preferred some form of highlighting, maybe slightly different font colours or background colour for each level, to show where the expansion starts and ends. I found when expanding an item, I wanted to read the addition but couldn’t quickly identify where it ended, which made it more difficult for my brain to selectively read just the expanded part.
I’m trying to find the best user experience for progressively expanded reading. Footnotes, for example, are only one indentation level and honestly most of them online are a pretty terrible experience to click on.
Footnotes on sidebars are good for most simple cases but insufficient for more elaborate structures, but you can in fact put a footnote in a footnote to get more than one indentation level.
I like the idea of footnotes on footnotes. I can see an interesting exploration in this space where footnotes (and footnotes of footnotes) appear as a bottom bar/appearing over the text. Sidebars don’t work IMO as on smaller screens there is not enough horizontal space, and mobile is where a lot of text is read.
a) I love Spaced Repetition systems mainly bc it feels empowering that you actually have a choice on what you will remember b) I cannot wait until there will be a fully AI driven one. I just want to chat via voice about concepts, and the agent attacks the question from different perspectives as opposed to asking in a statically formatted question-answer format c) I want that bc it will be nicer to just talk and bc the biggest problem with the system rn is that you memorize the static question–answer pair, association that is often tied to the context of your deck, and it’s often hard to use the knowledge in the real world. e.g.: I remember having in my deck a card that uses these words, asks in this specific way about something and I remember correctly that the answer is X. I think the only way I know out of this dynamic is to do more than one card per concept and to do reverse cards, but this is a lot of maintenance and in practice I rarely do it. If you have any pointers on how to deal with this please let me know!
There is another way, which is to try prototyping this out! The implementation seems pretty simple, but actually finetuning (disambiguated) it to be useful would take time I’m sure. But maybe it’s an upfront cost..
I asked Opus 4.8 to see what are solutions to this in the two texts i liked about Anki https://www.lesswrong.com/posts/7Q7DPSk4iGFJd8DRk/an-opinionated-guide-to-using-anki-correctly and https://augmentingcognition.com/ltm.html
The two additional solutions are:
– make cards very atomic (compare a more complex one-card-anki to four-cards-anki with the same knowledge, but a lot more atomic—atomic wins bc this type of knowledge is quicker to retrieve and more adaptable to the real life context, e.g.: in real life context you may need only 2 of 4 atomic bits, and you need to retrieve cards really quickly, and perhaps reassemble with different knowledge bits)
– to formulate cards from an angle of the real-world trigger you will most likely have.
-----
Claude Opus 4.8 (full-answer)
-----
Both texts address your exact problem (recall getting welded to the deck’s phrasing and context). Here is everything they offer, grouped by solution.
From “Guide on using Anki correctly”
Name the card for its real-world prompt, not its content. The core fix for “I only recall it when asked in that specific way.” Ask what real-life trigger should surface this knowledge, then build the card around that trigger. Their example: don’t name a card “length of whale intestine,” name it “animal fun facts,” because the real moment of retrieval is wanting an animal fun fact, not being quizzed on whale anatomy. If the natural real-life prompt genuinely needs a longer card, make the longer card.
Put the meaning in the answer, not just the bare token. The “24%” example: instead of “In 1950 Germany, what % of workforce was in agriculture? / 24%,” write “historical size of agriculture sector / 1950 Germany: 24% of workforce.” You are more likely to reach for “what do I know about the history of agriculture” than the narrow framing. Memorizing the isolated number without what it means is useless for real use.
Anchor to semantics, not visual shape. Their diagnosis of why cloze deletions fail is your exact problem: excessive context makes you learn “visual shape of the card leads to answer” instead of “semantic meaning leads to answer.” The problem is context volume, so strip context down until the recall path runs through meaning.
Redundancy as a deliberate feature. This generalizes your reverse-card instinct. Multiple cards with slightly varied prompts (small “epsilon” changes to wording and angle) train you to recognize the whole solution space rather than one frozen pairing. Cards should reinforce the pattern “circumstances lead to solution,” which is what makes knowledge fire in the real world.
Brevity forces transfer. Max ~9 words for most cards, at the absolute most 3 bullets / 18 words. If you cannot compress it, you do not understand it well enough to break it down. Short cards recall faster and resist context-anchoring.
From “Augmenting long-term memory” (Michael Nielsen)
Atomic questions you reassemble in unexpected ways. Break each idea into its smallest pieces. His soft-link example splits one routinely-missed card (“create a soft link”) into “what’s the basic command” and “what order do the arguments go.” The payoff he names directly: you later “assemble the atomic questions in an unexpected way.” Atomicity is what lets knowledge recombine in novel real situations instead of firing only as one block.
Never orphan questions, and add several per concept. This is the stronger version of your “more than one card per concept” instinct. He makes it a rule to never add just one question, always at least two, preferably three or more, so the fact becomes “the nucleus of a bit of useful knowledge.” Lonely orphan cards get missed constantly and are a waste. A tightly interconnected web of facts retains and transfers far better than isolated ones.
Multiple passes, 5 to 20 questions per source. Re-read hard material several times, adding questions after each pass. This layers the concept from different angles, which is a manual version of the “attack from different perspectives” agent you want.
No yes/no questions (from his work-in-progress notes). They let you pass on recognition without real recall.
I just created Dig Text, a new writing standard, which allows you to read the shortest version first, and dig deeper only where it interests you. I thought about dig about ten years ago, lol, and now, with the help of Claude Code, I spent two weeks thinking it through and coding it. It’s still in BETA, so I’m gathering feedback, but I would love if you try it and let me know what works and what doesn’t. https://digtext.github.io
I had a similar thought for math proofs, where you should be able to see an overall proof sketch first, then expand on any point to see the intuition/longer proof in a hierarchical way.
Related:
https://vlad.roam.garden/How-do-I-read-things-on-the-internet (LW crosspost). This is bullets points style, default expanded. I don’t like the indentations though, I think it is a bad idea.
Feedback:
IMO there should be a variant of text expansion where the original text gets replaced by the longer version.
The “Plus” icon is an improvement over having a new paragraph / indentation, but still too distracting, especially the purple color.
Keyboard navigation is a bit clunky.
Pressing Enter should move focus somewhere sensible, rather than eating the focus. Probably to the ending parenthesis for expansion, and to the original + icon for collapsion.
Maybe just me, I expect ArrowRight on the Plus icon to expand the text too.
Let me use ctrl+z ctrl+y to undo / redo
Would be nice to have a “default expand n layers deep” button.
Please show me an example of
((text))and:::dig:::tags in the demo.Double parentheses syntax
I like it
You are not allowing double parentheses to be escaped, try \((a))
Your parser is wrong, try ((a((b))c))
btw the selection of text in your demo is fucked (Firefox; Windows if that matters), it is selecting text one line lower than where my cursor is.
I am slightly annoyed that the Plus icon is not centered properly. It is a little bit too low.
The way your examples are structured with one sentence expansions seems bad. I don’t want to spend a few seconds clicking the button just to read one more sentence
It makes me want to have some sort of preview (full preview or how many levels / characters hidden)
Marginally related on the “dig deeper” part:
These pages uses side panels for Obsidian-style page linking. (via)
https://cq2.vercel.app/
https://notes.andymatuschak.org/About_these_notes
https://delve.a9.io/ (seems defunct now)
Thanks for your patience, I am back from my time off.
You listed a lot of bugs, visual problems that I want to fix. Thank you.
I agree that just text color for pluses may be a better choice here.
I agree that the best use is when digs don’t hold a small amount of text, but this is up to the writer
“Pressing Enter should move focus somewhere sensible, rather than eating the focus. Probably to the ending parenthesis for expansion, and to the original + icon for collapsion”. I think this is very interesting. I have been thinking how to approach this. I think there are tradeoffs here, mostly simplicity. There is a way, actually, to have preview text (see dig example and text in parenthesis). I will think if there is a sensible way to bake it in more universally into the dig format eg.: • {{preview text}} or similar.
What do you mean by preview text? Do you mean
Here is some text that is shown ((and some text that is hidden))? I don’t consider that as preview text, I am imagining text showing up on hover (unsure about exact design). If you mean something else, I don’t understand.Also I still have no idea how the :::dig syntax work.
I meant this bit on https://digtext.github.io (see image)
hover don’t work on mobile so its not a viable solution IMO
how would do u feel about the solution from the image?
I didn’t fully thought :::dig syntax through but I imagined adding a script on top of the site and inside dig tags all bulleted lists would render as dig text. Am I missing something in this solution?
I like the idea, though I also already really like classical options such as footnotes, and don’t even mind footnotes on footnotes.
One UI feedback point: when expanding an item, it is currently shown exactly like the main text. I’d have preferred some form of highlighting, maybe slightly different font colours or background colour for each level, to show where the expansion starts and ends. I found when expanding an item, I wanted to read the addition but couldn’t quickly identify where it ended, which made it more difficult for my brain to selectively read just the expanded part.
thank you for the feedback, noted
I used such three level bullet points style in https://www.lesswrong.com/posts/iBg6AAG72wqyosxAk/the-badness-of-death-in-different-metaethical-theories
How similar it it to an army of footnotes, collapsed sections or outright to early stages of a Wiki Walk?
I’m trying to find the best user experience for progressively expanded reading. Footnotes, for example, are only one indentation level and honestly most of them online are a pretty terrible experience to click on.
Footnotes on sidebars are good for most simple cases but insufficient for more elaborate structures, but you can in fact put a footnote in a footnote to get more than one indentation level.
I like the idea of footnotes on footnotes. I can see an interesting exploration in this space where footnotes (and footnotes of footnotes) appear as a bottom bar/appearing over the text. Sidebars don’t work IMO as on smaller screens there is not enough horizontal space, and mobile is where a lot of text is read.