Rubric beats prompt for cram flashcards. The reason is arithmetic, not pedagogy.

The claim is narrower than 'rubrics produce better cards'. The claim is that cramming changes which kind of bad card matters. In a calm study session, a bad card is a card you edit, and editing time is fine because you are learning while you edit. In a cram session, editing time is dead time: you are not retrieving, not closing knowledge gaps, not progressing through the deck. A prompt-only output that contains 30 to 40 bad cards in a 200-card deck (the rate measured against held-out evals at the field's average) is a 15 to 40 minute editing tax before drilling starts. A rubric-gated output drops the bad cards before emission, so the deck arrives drillable.

Four shapes specifically. (1) Stem references a fact the upload does not contain (pretrained drift). (2) Three short distractors next to one long correct answer (length-tell, you guess without reading the stem). (3) An option that says 'all of the above', 'none of the above', 'both A and B', or 'it depends' (filler-template, NBME flags this as an item-writing flaw). (4) The stem and the correct option do not parse together (singular-plural mismatch, article disagreement). Each one is solvable by a quality rubric run before emission. None of them is solvable by a better prompt alone, because the prompt cannot verify the upload, count characters across options, or grammar-parse stem-option pairs reliably enough to gate.

You can, and on a single short PDF on a calm afternoon it gets you most of the way. The mode of failure is at scale and under time pressure. The grammar and length checks transfer cleanly to a prompt: instruct the model to count characters and verify article agreement before emitting. The filler-template ban transfers cleanly: list forbidden options. Source-anchoring is the gate that does not transfer: ChatGPT cannot easily verify a cited slide span exists in your upload, so the source-anchoring check degrades to a self-report. The model will sometimes confabulate a citation. On a 200-card deck this means a residual ~5 to 10 percent bad-card rate even with a careful prompt, which is 10 to 20 cards of editing for a careful prompt vs zero for a rubric-gated generator. At 1 a.m. that is the difference between drilling 4 decks and drilling 5.

30 to 60 seconds per card is the post-hoc-review cost we use internally for the same four-criterion rubric, applied as a checklist rather than as in-flight gates. It is in line with the published cost of human item review on NBME-style materials (a careful reviewer with the source open in a second window finishes one MCQ per 45 to 90 seconds depending on item complexity). The math in this page uses 45 seconds as the midpoint. A 30-card editing tax at 45 seconds per card is 22.5 minutes; a 40-card tax is 30 minutes. Both numbers round to 'a deck of cramming time you do not get back'.

Nobody is throwing out the deck. The argument is about the tax to make the deck drillable, not about whether to use the cards that survive. The problem is that the bad cards and the good cards are interleaved, and there is no reliable way to tell them apart without reading every card with the source open. Skipping the editing pass and drilling the raw output means you internalize wrong facts (pretrained-drift cards that contradict your professor's slide), and you waste retrieval attempts on length-tell cards that you got right by reading the option list, not the stem. The interleaving is the cost. A rubric-gated generator removes the interleaving by emitting only the survivors.

Studyly scored 81.3 on a held-out three-document eval covering factual correctness, clarity, distractor quality, and question-type coverage. Unattle scored 78.0, Gauntlet 68.0, Turbolearn 57.8. The distractor-quality criterion has the widest tool-to-tool spread, which is the criterion most directly affected by an in-flight gate. The full leaderboard and methodology are at studyly.io/quality. The eval is not what makes the page argument true; it is the upper bound on how much editing a generator's output saves you. A 23.5-point spread on the same documents means the rubric-gated generator emits roughly 23 percent fewer flagged cards on average, which is the figure the editing-tax math uses.

Partly. The argument is true regardless of which rubric-gated generator you use. Any tool that runs the four in-flight gates honestly will save you the editing tax compared to a prompt-only output. Studyly is the version we built and benchmarked, so the numbers are ours. If you find another tool that scores comparably on the same held-out eval and runs the same gates, the argument applies to it equally. The argument breaks for tools that claim a rubric but apply it post-hoc as a UI checklist, because that just shifts the editing tax onto the student.

Both, but the channels are different. Editing time is the explicit channel measured here. Retention is the implicit channel: a deck without length-tell cards forces you to read the stem, which means you are practicing retrieval rather than recognition. The Karpicke and Roediger 2008 retrieval-vs-rereading retention gap (about 80 percent vs 34 percent one week later) is the upper bound on what retrieval practice can give you over passive review; bad cards that let you guess without retrieving leak the gap back. The cleaner the cards, the closer your effective retention sits to the retrieval ceiling.

The four-gate argument applies to any intake format. Source-anchoring degrades a little for scanned PDFs (OCR errors can prevent a clean span match) and for YouTube transcripts (timestamps replace slide numbers as the citation unit). The other three gates do not depend on the upload format. The editing-tax arithmetic is the same. A prompt-only output on a YouTube transcript still produces ~15 to 20 percent bad cards; a rubric-gated output still removes them.

Drop your remaining lecture decks and PDFs into a rubric-gated generator (Studyly is the one we built; the freemium tier on app.jungleai.com has no credit card gate). Skip the comparison shopping until the exam is over. Drill MCQs end-to-end on the first deck to get a baseline of what you do not know. Switch to free-response or case-style on the cards you missed, which forces cold recall instead of recognition. When you miss a card, click the source slide citation, read just that slide, then close it. Move to the next deck. The retrieval-with-feedback loop is the thing that produces the retention gap; everything else is overhead you should be removing.