There are hundreds of "study techniques." Cognitive science has repeatedly validated only three switches — the rest is packaging, including whatever your AI learning tool is selling you.
Day 53 argued that learning with AI requires guardrails. This issue hands you the three strongest rebars inside those guardrails. Dunlosky et al.'s heavily-cited 2013 review put ten mainstream study techniques on the table and rated each by strength of evidence — and most of them (highlighting, rereading, summarizing, imagery) scored low or context-dependent. Only "practice testing" and "distributed practice" earned high-utility ratings that generalize across ages and materials; interleaving earned a conditional "moderate." In other words, if you engineer learning as a system, there are really only three switches worth wiring in: retrieval, spacing, interleaving. This issue skips "what they are" (that's Psych 101) and covers three things that actually bite in practice: each switch's precise activation condition, its failure boundary, and a four-question checklist for selecting or building an AI learning tool — because almost none of them turn these on by default.
Roediger & Karpicke 2006 nailed it: after reading a passage, a "reread it" group and a "close the book and freely recall it" group looked similar at first — the reread group even felt more confident. But on the delayed test a week later, the retrieval-practice group won decisively. That's the testing effect: memory doesn't consolidate when you "see" material, it consolidates when you effortfully retrieve it. Dunlosky's review rated it top-tier, holding across ages and subjects. But it has two non-negotiable engineering conditions: retrieval must be effortful (free recall — not multiple choice you can guess, which tests recognition, not recall); and it must carry corrective feedback, or you "practice in" your wrong answers. And whether it worked can only be seen on a delayed post-test — an immediate test only reflects a short-term cache.
Turning any chatbot from an answer-machine into an examiner takes one system prompt:
# Retrieval loop: force recall first, then reveal + correct, then schedule a delayed post-test
SYSTEM: You are my retrieval-practice coach. Strict rules:
1. I name a topic I just studied; you pose 3 open-ended free-recall questions
(no multiple choice — that's guessable, tests recognition not recall)
2. Reveal nothing — no hints, no answers — until AFTER I answer
3. Then give the model answer + point-by-point correction (no feedback = errors get practiced in)
4. Attach a delayed post-test plan per question: re-quiz me on it in N days
Rules 2 and 3 are the crux: push the "reveal" to after I've answered, and force feedback — exactly the mechanism by which Day 53's "tutor version" guardrail erased the learning harm.
Cramming scores high on the spot and feels fluent, but decays fast; spread the same time across spaced sessions and long-term retention improves markedly. The real engineering question is how long the gap should be. Cepeda et al. 2008 — 1,350+ people, study gaps up to 3.5 months, final tests delayed up to 1 year — mapped a "ridgeline": final performance rises then falls with the gap (an inverted U), and the optimal gap grows with the test delay but shrinks as a proportion of it — roughly 20–40% of a 1-week delay, dropping to about 5–10% of a 1-year delay. A second counter-intuitive result comes from Karpicke & Roediger 2007: expanding schedules only win short-term; equal-interval wins long-term. What actually matters is that the first review is delayed and effortful, not whether later gaps expand.
# Set gaps as a FRACTION of the deadline, not fixed days
retention_interval = days_until_needed # days until interview / launch / exam
first_gap = 0.1 * retention_interval # first review ~10% of the deadline
# Cepeda 2008: optimal gap grows with the deadline, shrinks as a fraction of it
# Karpicke&Roediger 2007: use equal intervals long-term, don't fetishize expanding;
# what matters is the first review is delayed off the study session and effortful
Why does interleaving beat blocking? Rohrer & Taylor 2007's math experiments give a mechanism-level answer: with blocked practice you already know "this whole page uses the same formula," so you only train execution, never discrimination — and you flounder on a mixed test. Their error analysis showed the blocked group's mistakes were overwhelmingly "picked the wrong method" — not "couldn't compute," but "didn't know which tool to use." Interleaving shuffles problem types, forcing you to first judge "which kind is this, which approach applies" on every item — training exactly that discrimination step. But that also bounds it: interleaving's dividend comes from distinguishing confusable items, so it only helps material that is mutually confusable and requires strategy selection (derivatives vs. integrals, optimistic vs. pessimistic locks, several similar bug classes). For independent skills with no "wrong-method" problem, interleaving just adds context-switching cost. This is precisely why Dunlosky rated it only moderate — utility is highly material-dependent.
# Interleave only the "confusable pairs"; keep independent skills blocked
confusable = [("derivative","integral"), ("BFS","DFS"), ("optimistic-lock","pessimistic-lock")]
independent = ["learn SQL", "ear training"] # no discrimination problem -> only adds cost
for pair in confusable: # shuffle the two types together,
quiz = shuffle(problems(pair)) # forcing "which kind is this?" first
# When asking AI to generate: require mixed order, DON'T label the type
Collapse the first three sections into one selection criterion. Most AI tutors default to "explain first, then hand over the full answer" — which is rereading, and the very source of the learning harm in Day 53's PNAS "vanilla GPT" group. An even more common marketing line is "intelligently adapts to your learning style" (visual/auditory/…): Pashler et al.'s 2008 authoritative review states plainly that this "meshing hypothesis" lacks supporting evidence — tailoring instruction to a learning style does not improve outcomes. So "sells a learning style" isn't a feature, it's a pseudoscience signal. There's also an engineering reality: lab → classroom effect decay. The three switches show handsome effect sizes in the lab, but in real settings effects generally shrink and vary by material (Dunlosky's review keeps stressing "utility is conditional"). The implication is direct: outsourcing the discipline of the three switches to a tool's defaults is more reliable than sustaining it yourself daily — because micro-dosing and intermittent execution are exactly how these switches fail.
# Four questions for any AI learning tool (use it to select OR to build)
[ ] Force me to recall before revealing? No = rereading disguised as chat
[ ] Schedule a DELAYED post-test, not on-the-spot? No = tests short-term cache only
[ ] Mix confusable content in the quiz? No = discrimination training lost
[ ] Give corrective feedback every answer? No = wrong answers get practiced in
# Deduction signal: the pitch is "adapts to your learning style"
# -> meshing hypothesis has no evidence (Pashler 2008)
Passing all four means the three switches are actually set as defaults; otherwise the discipline still rests on you — and you won't sustain it.
① Retrieval — activates on: free recall (not multiple choice) + corrective feedback + a delayed post-test; fails on: guessable, no feedback, on-the-spot testing only.
② Spacing — activates on: first review delayed to "~10% of the deadline," multiple doses; fails on: spacing just once (micro-dose), stretched too long off the ridge, fetishizing expanding schedules.
③ Interleaving — activates on: mutually confusable material requiring strategy selection; fails on: interleaving unrelated independent skills (pure switching cost).
Core: there are hundreds of study techniques; only three reliably generalize across materials (Dunlosky 2013). Selecting an AI tool is one question — does it make these three its defaults, or is it selling "learning styles"?