How to Create a Useful Deck

FSRS is good at deciding when you should see a card again. It has no opinion at all about whether the card itself is any good. That part is still on you — and a deck of bad cards will waste your reviews no matter how well-tuned the scheduler behind them is. Here's what tends to separate a deck you actually learn from with one you just click through.

One idea per card

The most common way a deck goes wrong is cramming too much onto one card. A front that reads "Explain the causes of the French Revolution" isn't a flashcard — it's an essay question, and every time you see it you'll rate yourself on vibes rather than on whether you actually recalled something specific.

Split it. One card per fact, one card per definition, one card per step. If a "single" card keeps needing partial credit — "well, I remembered three of the five causes" — that's a sign it's really five cards wearing a trench coat. Smaller cards are also easier to write a real back for, which matters more than it sounds like (more on that below).

Write the back like an answer, not a summary

A good back face lets you know immediately, unambiguously, whether you got it right. A bad one is technically correct but too vague to self-grade against.

• Front: "What does FSRS stand for?" Back: "Free Spaced Repetition Scheduler." Either you knew it or you didn't.
• Front: "What happened in 1789?" Back: a paragraph. You'll always find a way to convince yourself you "basically" knew it.

If you're using Written or Sentence builder quiz mode, this matters even more — those modes grade you by comparing what you typed or arranged against the back face, word for word (punctuation and case aside). A back face written as a clean, short answer is one you can actually type from memory. A back face written as a paragraph isn't.

Use both faces deliberately

Front and back both support rich text — bold, lists, headings, code — and you can embed images, audio, and video inline, not as a separate attachment. A few ways that's actually useful, rather than just decoration:

• A vocabulary card with a native-speaker audio clip on the front and the translation on the back, so you're training listening comprehension, not just reading.
• A diagram embedded on the front of an anatomy or circuit card, where a text description would just be worse.
• A formula written with $$…$$ (block) or \(…\) (inline) LaTeX, so it's typeset properly instead of sitting there as raw notation.

There's also a third, optional field — the note — that never gets studied and never counts toward scheduling. It's for you: a mnemonic, a source link, a reminder of why you added the card. It stays private to you even if you publish the deck or someone forks it.

Split decks by what you're actually testing

A deck isn't a folder — it's a review pool. Every card in a deck gets mixed together during Study, Quiz, and Browse, so a deck that mixes two unrelated subjects means your reviews for one keep interrupting your focus on the other. If you notice yourself thinking "I only want to review the verb conjugations right now, not the whole language deck," that's usually the signal to split it into a separate deck rather than trying to filter within one.

The reverse mistake also happens: decks so narrow (five cards) that FSRS never gets a large enough pool to schedule meaningfully. There's no hard rule here — but if a deck fits comfortably as "the thing I'd search for in Discover," it's probably sized about right.

Let AI generation write the first draft, not the final one

If typing every card by hand is the barrier, AI generation gives you somewhere to start rather than a blank editor. It takes a PDF, an image (a photo of your notes works fine), a DOCX or PPTX, a short topic description, or a recorded lecture, and turns it into a set of card suggestions you review before anything is saved to your deck.

It's worth treating the output as a draft, not the finished product. The model doesn't know which five facts you actually need drilled versus which ones you already have cold — you do. The review screen exists so you can trim it down to what's actually worth your review time, tighten a back face that's too wordy to self-grade against, and merge anything that's really testing the same idea twice.

The short version

Split until each card tests one thing. Write backs you can grade yourself against at a glance. Use rich text and embeds where they genuinely help recall, not as decoration. Group cards by what you want to review together, not by what you'd file together. And if AI generation gets you started, finish the job yourself — the five minutes you spend editing suggestions is where a mediocre deck becomes a good one.