Anki is a flashcard app built on spaced repetition. Rather than reviewing on a fixed schedule, it tracks how well you remember each card and shows it to you at exactly the point you're likely to forget it. Common uses are language learning and medicine.
What puts Anki above most alternatives (Wanikani, JPDB, etc.) is the Free Spaced Repetition Scheduler (FSRS). Unlike the older SM-2 algorithm, FSRS adapts to your actual memory rather than using population-average parameters. The difference compounds over time — your intervals become genuinely personalised to you.
This is my current setup. Most of it applies to any Anki user, but the card creation section is skewed towards Japanese.
Preferences
Next Day Starts At
The problem this setting solves: if you do reviews from 23:30 to 00:05, you'll have completed two different days' reviews in a single session. Cards with short intervals will appear twice, and you'll remember them much better than expected, corrupting the scheduling data.
The standard fix is to set this to a time when you're always asleep — later than midnight if you're a night owl.
An alternative I've ended up using: keep it at midnight and do reviews just after the reset. This means you're doing the next day's reviews. The tradeoff:
Advantages:
- Leeway to skip a day without reviews accumulating.
- No pressure to finish before bed.
Disadvantages:
- Skipping a day means catching up with two days of reviews at some point.
- You can't start reviews before midnight.
- Long learning/relearning steps become problematic — you need to finish before sleeping to avoid the double-review problem.
I fell into this schedule accidentally. It works for me now, but I wouldn't deliberately recommend it.
Learn Ahead Limit
I used to set this to a very high number so I was never left waiting for a learning card. I've changed my position on this.
A high learn ahead limit makes it easy to spam "again" at the end of your session rather than stopping. This drives up card difficulty artificially, and seems to degrade FSRS optimisation — several people in Japanese learning communities have swapped back to default parameters after cutting their learn ahead limit, getting equivalent retention with less time spent.
My current recommendation: set this to 0, or at least well below your shortest interval.
Show Next Review Time Above Answer Buttons
If there's any chance the interval shown will influence which button you click, turn this off. There's no benefit to keeping it on.
Deck Settings
New Cards/Day
This is the primary lever for controlling long-term workload. A rough rule of thumb: expect 7–10 reviews per day for every new card you add daily. Plan accordingly.
If you're cramming before a deadline, drop your FSRS desired retention as low as you can tolerate and increase new cards to cover the material faster.
My current approach for Japanese is to set new cards/day to 9999 and add cards from sentence mining as I encounter them. This creates inconsistency — a hard book might give 100 new cards in a day, an easy one barely 10. It suits how I study but requires comfort with variable daily load.
Maximum Reviews/Day
Leave this at 9999. Capping it doesn't reduce your workload in any meaningful sense — it just means you forget the cards you didn't review. The only way to actually reduce load is to do fewer new cards.
Learning Steps
There's no conclusive research on optimal learning steps. Most people use a single step somewhere in the 3–20 minute range.
One approach worth knowing: some learners skip learning steps entirely, instead burying any new card they fail immediately rather than clicking "again". This lets them use the first "again" click as a difficulty signal rather than a failure marker. Whether that's worth it depends on your workflow.
FSRS-7 may eventually include a short-term scheduler, which would make this setting largely moot. Until then, a single step in the 5–10 minute range is a reasonable default.
Relearning Steps
Same lack of research as learning steps. At least one relearning step is generally recommended; more than one isn't considered necessary.
Longer steps are thought to be more effective short-term, but they should comfortably finish within the same day — steps that carry over mess with FSRS scheduling. Most people use 3–20 minutes. One Japanese learner I follow uses 45 seconds, reasoning that long-term memory consolidation starts around 30 seconds with adequate distraction between.
Leeches
A leech is a card you keep failing. I used to just power through them. I've stopped doing that.
Two problems with ignoring leeches:
- You end up spending a disproportionate amount of time on them — some I was reviewing every other day for weeks.
- Including leeches when optimising FSRS parameters can make your memory appear worse than it is, increasing intervals across your whole deck — not just the leeches.
My current approach:
- Unimportant card? Delete it.
- Important but badly written? Rewrite it.
- Important, can't rewrite, but memorable? Add a mnemonic.
- Vocabulary card where none of the above work? Move it to a separate leeches deck.
For the leeches deck, I don't require myself to review it daily — I do it when I feel like it. This reduces time spent on difficult cards and helps separate pairs I was confusing. I plan to reintroduce cards to the main deck when their interval reaches around a month.
One limitation: Anki doesn't have a built-in way to reset a card's lapse count, so removing the leech tag just results in it being re-added quickly. Worth keeping in mind.
Display Order
New/Review Order
Show after reviews is what I currently use. It means you don't learn new cards if you fail to finish your reviews — optimal. The downside (easy "again" spamming at the end) is neutralised by keeping the learn ahead limit at 0.
Show before reviews prevents again-spamming even with a learn ahead limit, but makes the session feel harder from the start and risks adding new cards you then can't consolidate with reviews.
Mix with reviews spreads the difficulty throughout the session. Reasonable option.
Review Sort Order
If you're finishing all your reviews every day, this doesn't matter.
If you sometimes don't finish: descending retrievability maximises retention on the cards you do review, at the cost of completely ignoring some cards. Ascending retrievability gives broader coverage at lower overall retention. Pick based on whether you'd rather have high retention on a subset or moderate retention across everything.
FSRS
Turn this on. There is no reason to use Anki without it.
Desired Retention
This is the second way to control daily workload, alongside new cards/day.
Some rough guidelines:
- Above 97%: the review count increase is enormous for minimal real-world benefit. Don't do this.
- Above 94%: sharply diminishing returns. The workload is disproportionate to the retention gain.
- Below 70%: getting a large proportion of reviews wrong is demoralising and likely counterproductive.
A target of 85–90% is a reasonable starting point for most use cases. Recent Anki beta builds include tools to help you choose based on your actual data — worth using if you have enough review history.
Optimising Parameters
Make your optimisation filter as specific as possible. Different card types and decks have different memory characteristics — lumping them together produces parameters that fit nothing well.
Other useful practices:
- Exclude reviews before a date when your technique or card format changed significantly.
- Exclude leeches.
- Exclude suspended cards.
- You can also optimise your leeches deck using parameters derived from your main deck.
Auto Advance
Reveals the answer automatically after a set time and can fail the card if no response is given. Useful if you want to enforce a response-time limit. I used it for a while at 3.8 seconds but stopped once I was comfortable reviewing at speed naturally.
Burying
Burying postpones a card until the next day. I use it in three situations:
- Accidentally revealing the answer before attempting the card.
- Separating related cards that would give each other away if seen in the same session.
- Postponing a new card I haven't learned yet, rather than immediately clicking "again".
Add-ons
Note: add-ons are desktop-only. Avoid setups that require them to function correctly.
Generally Useful
FSRS Helper — Extends FSRS with extra tooling. Worth having if you're doing anything non-trivial with your parameters.
Advanced Browser — Lets you sort the card browser by note fields, first/last review date, and other stats. Useful for deck management.
For Sentence Mining
Anki Connect — Exposes an API so external tools (Yomitan, etc.) can create cards programmatically. Essential for any mining setup.
FieldReporter — Two features: reordering cards by a field (enables frequency sorting so you learn common words first), and backfilling frequency data into existing cards. Frequency sorting is essentially free gains — I'd consider it a must-have for vocabulary learning.
Cosmetic
Kanji Grid — Displays a grid of kanji/hanzi you've encountered or learned, sortable by JLPT level, school grade, etc.
Review Heatmap — GitHub-style contribution graph for your review history. Not essential, but a nice home screen addition.
Anki Leaderboard — Adds a global leaderboard. More motivating than I expected.
Reviewing
Grading Cards
Simple approach: only use "again" and "good". Forgot it → again. Remembered it → good. Some people add an add-on to remove the other two buttons entirely. This is fine and FSRS adapts to it.
More precise approach: use all four buttons based on recall effort. The critical rule: "again" is the only failing grade. If you remembered it at all — even slowly, even with effort — use "good" or "hard". Reserve "again" strictly for genuine failure to recall.
"Easy" is for cards that were effortless — usually ones where the interval should grow faster. Use it accurately if you can, but FSRS will compensate if your calibration is imperfect (e.g. by making "hard" produce the same interval as "good" if it observes that pattern in your data).
Creating Cards
Sentence Mining
For language learning, sentence mining is the process of turning words and sentences you encounter in native material into Anki cards. Tools like Yomitan (a browser dictionary extension) integrate with Anki Connect to generate cards with a single keypress — reading, definition, audio, and sentence context included. For more detail on the workflow, this deserves its own post.
The Generation Effect
The generation effect is the finding that you retain information better when you produce it than when you passively read it. For Anki, this means cards that require you to generate the answer are more effective than cards where you read it.
For language learning, this suggests NL → TL (native language to target language) cards are theoretically better than TL → NL. I'm not sure the generation effect outweighs the increased difficulty in practice, but it's worth thinking about card direction deliberately.
My current plan is to use monolingual cards: a definition on the front, the term on the back. The definition should uniquely identify the term where possible, supplemented by example sentences or hints if not. For terms where I can't yet understand a monolingual definition, I fall back to TL → NL.
Writing has also been shown to aid retention. For Japanese/Chinese specifically, writing out characters for new vocabulary is worth considering — but selectively. Common components like 物 or 者 appear in dozens of words. Writing them once or twice across the words they appear in is sufficient; writing them for every word is a waste of time.
Cloze Deletion
Cloze deletion removes a word or phrase from a passage and asks you to recall it: ________ is the anaerobic catabolism of glucose.
The main advantage over traditional Q&A cards is the ability to generate multiple cards from one note. For example:
In a kanji compound, ツ becomes ッ when it precedes a character from the
{{c1::カ行}}, {{c2::サ行}}, {{c3::タ行}}, or {{c4::ハ行}}.
This generates four separate cards from a single note — each one testing a different blank. It's by far the easiest way to turn arbitrary text into reviewable material.
Miscellaneous
Queries
Anki's search syntax works everywhere: the card browser, filtered decks, FSRS optimisation, and certain add-ons. Worth learning the basics — is:new, is:due, prop:ivl>=X, tag:leech, etc.
Filtered Decks
Filtered decks pull cards matching a search query into a temporary deck for review outside the normal schedule. You control whether Anki reschedules based on your answers. Useful for:
- Previewing new cards without affecting scheduling.
- Cramming before a test.
- Reviewing tomorrow's cards a day early.
- Reviewing only a specific tag or subdeck.
- Combining multiple decks for a single session.
- Working through your leeches in isolation.
Catching Up on a Backlog
Sort by descending retrievability and do as many cards as you can each day. This prioritises the cards most at risk of being forgotten while you rebuild momentum.
Doing Extra Reviews Ahead of Schedule
Create a filtered deck with the query prop:due<=X deck:Y where X is how many days ahead you want to review and Y is the deck name. A few important caveats:
- Enable "reschedule cards based on answers" — this is non-negotiable.
- Add
prop:ivl>=5(or higher) to avoid pulling in cards with very short intervals. Reviewing these early disrupts the scheduling more than the extra practice is worth.
FSRS Visualiser
https://open-spaced-repetition.github.io/anki_fsrs_visualizer/ — input your FSRS parameters and see projected intervals after different answer streaks. Good for understanding how your current parameters behave before committing to them.