Stats
Actions
Tags
From ai_tutor
Create, list, update, or archive Mochi spaced-repetition flashcards via the Mochi MCP server, AND the canonical reference for how Harrison wants cards designed (six commitments, templates, anti-patterns, self-check). Use when Harrison says "make a mochi card", "add this to mochi", "create flashcards for X", "push to mochi", "list my mochi decks", or any flashcard/spaced-repetition task targeting app.mochi.cards. Also load this whenever drafting cards anywhere — it is the design canon.
How this skill is triggered — by the user, by Claude, or both
Slash command
/ai_tutor:mochiThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Spaced-repetition flashcard system at app.mochi.cards. Use the MCP tools (prefix `mcp__mochi__*`), not the API directly.
Spaced-repetition flashcard system at app.mochi.cards. Use the MCP tools (prefix mcp__mochi__*), not the API directly.
This skill has two halves:
mcp__mochi__list_decks if mismatched)gMybxR6k — important ap (AP / college-board content; AP Stats, AP CS, etc.)uRirM1bJ — Flashcards (default / misc)H93OQZZU — Vocabulary Wordsv3n8iBPK — baron 500, vxGo86k1 — baron 1500 (SAT vocab)LmeXZY7P — cool facts\n---\n on its own empty line (blank → --- → blank)## Title on the front so the question side stays short; body on backtags array, not in the card bodyMochi enforces 1 concurrent request per account server-side. Batch create_flashcard calls must serialize — firing two MCP tool calls in parallel gets ECONNRESET or 429.
ECONNRESET — usually transientNo real-time. Any "react to Mochi changes" workflow has to poll.
op item edit nleaoscl6pppjplhuibj6yf2yu "api_key[password]=NEW"
Then re-register the MCP — Claude stores env vars literally in ~/.claude.json, doesn't resolve op:// refs at runtime.
Mochi API access requires a Pro subscription.
Every card pushed to Mochi must be concise, interest-centered, insightful, engaging, accurate, and corroborated. If a draft fails any of those, do not push.
Concise and direct. Answer ≤15 words for the load-bearing fact. Front question is one clause. No "it is important to note," no padding, no half-hearted phrasing. If the back needs a paragraph, the card is wrong — split it.
Centered in interest. Card content comes from sources Wozniak and Nielsen would recognize: textbooks, primary papers, encyclopedias, Wikipedia. Never from corporate blog posts about that company's own product. Vendor-marketing prose is a card-source ban. (Wozniak Rule 20 enumerates books/journals/encyclopedias/dictionaries; blogs are absent. Nielsen: "if you Ankify misleading work… you're actively making yourself stupider.")
Insight or breadth, never trivia. A card earns its place by capturing a non-obvious "aha" OR by compressing a wide territory into a retrievable summary. Symbol-name lookups ("what does ω mean?") and verbatim-quote regurgitation fail this bar.
Engaging prose. Voice over varnish. The card should read like a sharp human noticed something, not like an encyclopedia summary. Cut padding, keep the anchor — one concrete person, scene, or analogy that swaps in for the abstraction. Matuschak: "the most important thing to optimize is the emotional connection to your review sessions." Boring cards are diagnostic of bad cards.
Correct and current. Every claim cross-checked against ≥1 source besides the originating session. Volatile content (software versions, recent research, products) is date-stamped in the prompt ("as of 2026-05") or attributed ("Jones 2011 claims…"). Wozniak Rule 19: date your knowledge.
Corroborated. Settled definitions: 1 authoritative source. Active research: ≥2 independent sources. Product/tool recommendations: ≥3 independent reviews or strong consensus. If you can't corroborate, the card waits.
The load-bearing rules from the canon. Every check below derives from them.
Minimum Information Principle (Wozniak Rule 4). "The material you learn must be formulated in as simple way as it is possible." One retrieval target per card. The principle is about per-card simplicity, not per-fact uniqueness — multiple cards on the same fact (different angles) is encouraged.
Do not learn if you do not understand (Wozniak Rule 1). Cards capture insight; insight requires comprehension first. Never card a fact the learner has not actually grasped.
Combat interference (Wozniak Rule 11). "If knowledge of one item makes it harder to remember another item, we have a case of memory interference." Wozniak names interference the single greatest cause of forgetting. Near-twin cards (chain vs product rule, t vs z, hoop vs sphere) need disambiguating cues baked into the front.
Avoid sets and enumerations (Wozniak Rules 9–10). "Name all members of X" cards fail because the brain enumerates differently each rep, fragmenting the trace. Convert to per-member cards with member-specific cues.
Five properties of a good prompt (Matuschak). Each prompt should be:
No orphans (Nielsen). "5–20 cards or don't bother." A single isolated card on a topic does not stick. Either build out 2–3 connected cards or drop the singleton.
Multiple cues for one fact (Nielsen). Forward, backward, applied — three angles on the same load-bearing fact, not three reformulations of the front.
Generation > recognition (Slamecka & Graf 1978; Roediger & Karpicke 2006). The learner must produce the answer before flipping. Recognition-only cards forfeit most of the testing-effect benefit.
Use in roughly this priority order.
When: capturing "when problem says X, your move is Y." Most valuable for AP-exam content. Front: "When you see [situation/feature], what do you reach for?" Back: named technique or first move, ≤10 words.
Example (AP Stats): Q: When a problem describes "before and after" measurements on the same units, what's the first move? A: Compute differences within unit, then 1-sample t on the differences. (Paired, not 2-sample.)
When: two concepts/techniques are easy to confuse (chain vs product rule, paired vs independent, hoop vs sphere). Without these cards, interference produces leeches. Front: "X vs Y: what's the operational test to tell them apart?" Back: the discriminating cue (not both definitions).
Example: Q: Paired vs independent samples — what's the operational test? A: Can you randomly shuffle one group's data without losing information? Yes → independent. No → paired.
When: capturing physical/conceptual intuition behind a result. Back must be a mechanism, not a formula. Front: "Why does X happen?" or "Why is X true mechanistically?" Back: a causal sentence. If the back is an equation, this template is wrong; use T5 instead.
When: a theorem/test has N conditions. One card per condition, plus optionally one count card. Never pack. Front: "For [theorem] to apply, what condition involves [aspect]?" Back: the single condition.
When: carding a formula. Carding the symbol-string alone is trivia. Pair it with: a trigger card (T1), a symbol-meaning card per symbol, and a limiting-case sanity card. Bad pattern: Q: "Formula for kinetic energy?" A: "½mv²" Good split: trigger card ("when you need translational KE…"), symbol cards (m, v), limiting-case card ("KE at rest?" → 0).
When: a load-bearing claim deserves over-carding. Make 3 cards on the same fact:
When: the claim is research-current, contested, or surprising enough that who said it matters. Prevents memorizing the claim as ground truth. Front: "What does [Author Year] claim about X?" Back: the claim, not as a free-floating fact.
When: the content has consistent ordering (a formula, a famous quote, a sequence step). Cloze preserves the structural frame and tests the missing token. Avoid cloze for non-ordered content (sets, lists of unordered properties).
Example (Alan Kay quote): "if the personal computer is truly a {{new medium}} then the use of it would actually change the {{thought patterns}} of an {{entire civilization}}, {{Alan Kay}}, {{1989}}"
When: the learner just made a specific mistake worth not making twice. Front: "What's the most common mistake when using X?" Back: the mistake + the symptom that catches it.
When: a problem type has a characteristic opening move that, once known, unlocks the rest. Front: "Given a problem with [feature], what's the first move?" Back: the opening step + which feature triggered it.
If a draft pattern-matches one of these, refuse-or-rewrite. Refusing is a positive action.
| Anti-pattern | Why it fails | Fix |
|---|---|---|
| Overlong answer (>15 words) | Answer becomes a guessing game; partial recall ambiguous | Split into atomic cards; the back is the punchline only |
| Set / enumeration ("name all 7 X") | Each rep produces different orders, fragmenting trace | Per-member card with member-specific cue, OR cloze each member in a meaningful frame |
| Ambiguous prompt ("What does GRE stand for?") | Multiple correct answers depending on context | Prefix domain context inside the prompt |
| Yes/no trap | 50% guess rate; allows pattern-match without retrieval | Convert to open question that demands the same insight |
| Trivia / encoded-not-understood | Passes review, no transfer | Pair with trigger card (T1) so the fact has a deployment context |
| Orphan card | One isolated fact, no neighborhood | Either build 2–3 connected cards or drop |
| Cards from inside the book | Depends on book's framing/example | Rewrite so a stranger could answer |
| Verbatim recitation | Memorizes phrasing, not concept | Card the underlying idea; if quote matters, cloze the load-bearing token |
| Symbol-name lookup ("What is ω?") | Trivia, no deployment cue | Replace with trigger ("When do you need ω?") + meaning card |
| Compound back ("…and…") | Two retrieval targets disguised as one | Split at the "and"; this is the single most common atomicity failure |
| Numerical answer ("What's the speed?") | The number is incidental; the setup choice is what matters | Card the setup decision and the failure mode, never the number |
| AI-slop encyclopedia tone | Generic textbook voice, no anchor | Rewrite in tutor voice with one concrete hook |
Decision tree, in order:
Wikipedia-tier sourcing as Wozniak frames it (verbatim): "the recommended source of basic incremental reading materials" — encyclopedic prose where "each sentence has a contribution to your knowledge."
Refuse, with a named rule, when:
The cost of a refused card is zero. The cost of a slop card pushed to Mochi is reviewing it for years.
Run this checklist on every drafted card. If any fails, revise or refuse.
If all pass: push, log, move on.
Card draft (insight: <what triggered it>)
deck: <domain>
tags: <concept-slug>
front: <prompt — prefer learner's verbatim phrasing>
back: <atomic, one fact, ≤15 words>
Approve / edit / skip?
Refuse with a named rule. Never just "no." If declining: "This is a Rule 1 case — not yet understood," or "Orphan card — only this concept; want me to draft 2 more siblings?" or "Trivia trap — symbol lookup with no trigger; want a trigger card instead?"
Push immediately on approval. No cooling-off. The SRS surfaces bad cards itself via leech behavior; that's faster than gut-check.
Pacing. Working ceiling 3–5 cards per hour-long tutoring session. 0 is fine when no insight surfaced. Quantum Country's 27/hr rate is for dense reading, not tutoring.
Wozniak — Dead Sea (the canonical Minimum Information Principle demo):
BAD:
Q: What is the Dead Sea? A: A salt lake located on the border between Israel and Jordan. Its surface and shores are 423 metres below sea level, Earth's lowest elevation on land. It is 377 m deep, the deepest hypersaline lake in the world. […]
GOOD (atomic split):
Q: Where is the Dead Sea located? / A: on the border between Israel and Jordan Q: What is the lowest point on Earth's land? / A: shores of the Dead Sea (423 m below sea level) Q: How deep is the Dead Sea? / A: 377 m
Matuschak — chicken stock (avoiding binary prompts):
BAD: "Does chicken stock make vegetable dishes taste like chicken?" (yes/no trap) GOOD: "How does chicken stock affect the flavor of vegetable dishes?" / "It makes them taste more 'complete'."
Matuschak — omelette (eliminating ambiguity):
BAD: "What's the first step to cook an omelette?" (multiple defensible answers) GOOD: "When making an omelette, how must the pan be prepared before you add the eggs?"
Nielsen — soft link (atomicity by failure-dimension split):
BAD: "How do you create a Unix soft link?" (two retrieval targets: command and arg order) GOOD:
Q: What's the basic command and option to create a Unix soft link? / A:
ln -s …Q: When creating a Unix soft link, in what order dolinknameandfilenamego? / A:filename linkname
Trigger + atomicity (AP Physics, rolling):
BAD (fused):
Q: Two solid spheres of different mass and radius roll without slipping down the same incline from the same height. Which arrives at the bottom faster? A: Same speed. v = √(2gh / (1 + I/(mR²))). For any solid sphere I/(mR²) = 2/5 regardless of m or R. Mass and radius cancel; only shape and h matter.
GOOD (split):
Q: Two solid spheres, different m and R, same incline, same h, rolling without slipping — which wins? / A: Tie. Q: Why doesn't mass or radius affect rolling-without-slipping speed for the same shape? / A: $v = \sqrt{2gh / (1 + I/(mR^2))}$ depends only on shape; for a solid sphere $I/(mR^2) = 2/5$ regardless of m, R. Q: For rolling without slipping, what dimensionless quantity decides who wins down an incline? / A: $I/(mR^2)$ — smaller wins.
Conditions decomposed (AP Stats, 2-sample t):
BAD: "What are the conditions for the 2-sample t-test?" (set card) GOOD (one card per condition):
Q: For 2-sample t, what condition involves how the data was collected? / A: Independent random samples (or random assignment for an experiment). Q: For 2-sample t, what condition involves the two samples' relationship? / A: Independence between samples. Q: For 2-sample t, what condition involves shape? / A: Each sample roughly normal OR n large enough for CLT.
{{...}}, not Anki's {{c1::...}}. Numbered groups: {{1::...}} {{2::...}} reveal together.$x$, block $$...$$. KaTeX renders most LaTeX; \dfrac, \hat{}, \bar{}, \text{}, \vec{} all work. gather*, multline, and \newcommand per-card don't — use Mochi's global LaTeX macros for repeated definitions (\hat{p}, \E, \Var).$$...$$ for the equation that is the answer. Cards wrapped entirely in display math with \text{} English are the canonical anti-pattern.\vec{v}, use that, not \mathbf{v}. Retrieval is surface-form sensitive; don't invent house style.\,\mathrm{...}.$ triggers math mode (escape \$); blank lines inside $$ break the block; markdown does not render inside HTML tags.flashcards skill ships an L1-recall / L2-understanding / L3-boundaries framework. Different lineage from Wozniak/Matuschak/Nielsen above; combine if useful, but push via mcp__mochi__* (skip its Anki workflow).Primary canon:
Cognitive psych anchors:
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub harqian/claude-tutor --plugin ai_tutor