Flashcard Creation Skill

For private academic quiz archives where the user has explicitly confirmed that a checked option is correct, prefer a short - explanation: bullet with clozes over a bare answer-only card. Cloze the decisive method, condition, contrast, or reason that makes the selected answer correct; do not merely hide the same final option twice. If the question or answer depends on an embedded figure, still add the cloze-rich explanation instead of treating the image itself as sufficient review support. If the Solution: line contains the embedded image, keep at least one cloze on that Solution: line too so the chosen option or structural descriptor is directly quizzable there.

For public academic quiz hints written as ::@:: cards, keep the hints in the same order as the archived/private question order. If the user wants slightly more context, add it mainly on the left-hand side prompt itself: use option-family cues, mutated-but-equivalent givens, or the decisive spectral, timing, or topology landmarks without copying the official choices verbatim.

For academic topic notes, do not limit cards to isolated definitions. Prefer a balanced mix of comparison cards, intuition cards, example cards, counterexample cards, and worked-example cards when the material supports them. If a section needs only a small clarification, enhance the existing flashcards; if it needs a substantial new cluster of distinctions or examples, add new flashcards instead of overloading one old card. For worked examples, place all required givens, formulas, assumptions, and numeric input data on the left-hand side before ::@:: so the card is fully answerable in isolation.

For mathematically technical academic notes, preserve the derivation or proof spine in the flashcards instead of testing only the final formula. A strong default is to add at least one card for the governing equation or setup, one card for the decisive derivation step or inequality, and one card for the final result or interpretation when the source material supports that structure.

When a course note is organized by topic pages and lecture weeks rather than stored chapter pages, avoid chapter-number-only prompts such as Chapter 2 / .... Use self-contained concept wording or the actual topic-note context instead.

For conceptual math-law cards in academic notes, a descriptive prompt is often better than forcing the formula itself onto the left-hand side. When such a card is genuinely conceptual rather than computational, prefer the descriptive prompt and, if needed, attach a targeted inline suppression comment on the same line instead of warping the card into a fake calculation prompt.

When to use

Be careful when users add inline comments or annotations such as <!-- check: ignore-line[...]: equation on left -->. Such comments are rationales for validator suppressions and not editing instructions. Never split or reflow the card simply because a comment mentions “left” or “right”; the left-hand portion always means the text that appears before the ::@:: or :@: separator, and it may be long when calculations are involved.

Invoke the skill when the user asks to “add flashcards”, “cloze this”, “quizify”, or similar. The target file must already exist; never create new files or edit submodules/private content without permission. Process one paragraph or logical block at a time and display the original text for confirmation. Aim for at least 92 % coverage of each paragraph and about 80–92 % of every sentence. Numeric facts or simple assignments may be clozed as atomic units.

Workflow

1. Consult examples. Scan the index of representative patterns earlier in this file for descriptions matching the current passage. Refer to the matching example numbers in your reasoning (e.g., “behaviour like example 3’s QA rewrite”).
2. Select a paragraph or contiguous block. Work on one at a time. Use a separate tool call per paragraph rather than batching. Pause after each block to retrieve the next.
3. Break text into meaning units. Look for definitions, names, dates, formulas, list items, pros/cons, conditional or contrast clauses, and semicolon‑separated ideas. Treat each as a potential card. If a sentence is mostly visible, add another cloze. Do not hide an entire short sentence with no anchor word; leave a hint word or two visible.
   • Handle semicolon‑separated clauses like a list with separate clozes.
   • For conditional connectors (if, when, once, unless), hide the text after the connector unless the condition itself is being tested.
   • For contrastive conjunctions (but, however, yet, etc.) make separate clozes on either side.
4. Insert clozes guided by examples. Use the matched examples as templates:
   • Wrap material based on meaning and recall effort rather than rigid word counts. Dense or confusing ideas get smaller clozes; obvious concepts can be larger.
   • If text already has formatting (bold, italic, code, underline, etc.), enclose the entire formatted phrase inside the cloze rather than inserting braces within it. For example use {@{__foo__}@} instead of __{@{foo}@}__.
   • Cloze both a leading proposition and its supporting detail if both are worth testing. Leave one or two words outside the braces for a contextual hint.
   • When several closely related noun phrases appear, split them into multiple smaller clozes instead of a single huge one.
   • Place the cloze where it best prompts recall—the first word, last word, or middle as the example shows.
   • Never break a LaTeX equation. Hide the entire $…$ or $$…$$ block as a single atomic cloze.
   • By default suggest inline clozes; only rewrite to ::@:: or :@: if the user explicitly requests QA style or if an example clearly uses that format.
   • Preserve Markdown, KaTeX, links, and line ordering. Lists are handled item‑by‑item unless an example demonstrates a combined deletion.
5. Apply the edits directly in the file and return the modified text. Provide alternate versions or explanations only if the user asks.

Continuous improvement

• Keep durable flashcard lessons in this document (and, for course notes under special/academia, in the academic-notes skill) rather than maintaining a separate sidecar learning log that drifts out of sync.
• After suggesting clozes, ask “which deletions are wrong?” or “would you prefer Q/A instead of inline cloze?” Use corrections to expand the examples or add rules. When the user states “this is my style”, treat it as a high‑priority rule and immediately encode it in this document, noting which examples or heuristics were updated.
• Analyse user edits using git --no-pager diff --word-diff --no-color path/to/file.md. Word diffs reveal boundary shifts and merged or expanded clozes; use them as training data and add corresponding examples.
• Watch for feedback about excessive fragmentation. Merge adjacent tiny deletions in future and update the rules accordingly.
• When the user modifies or adds contiguous paragraphs, capture the whole block as a single example tagged appropriately so the agent can recognise similar multi‑paragraph patterns later.
• Never suggest running any external command during editing. Command advice is irrelevant to this skill.
• If session memory grows large (≈20 examples or after a long session), summarise learned rules and cement them in this document. Prompt the user before consolidating.

Representative examples

Several patterns recur frequently; the first five entries below illustrate the kinds of transformations you should perform. When editing a new passage, recall which example best matches and mimic its cloze placement and style.

1. Dense math paragraph – multiple expressions and annotations require many inline clozes so each formula or qualifier can be tested separately.
2. Long technical paragraph – break a single sentence containing several logical ideas into separate deletions rather than one giant blank.
3. Circuit cancellation description – sequential procedural text with paired concepts gets split at each step; retain anchors such as “easiest way” or “however”.
4. Transform hierarchy statement – simple declarative sentence with nested subjects encourages small, discrete clozes for each noun phrase.
5. Concentration/spread of Fourier transform – comparative statement best handled by clozing each contrasted clause individually.

Below are the full before/after examples for reference (copy them into the agent’s reasoning when needed):

Input: If _f_ is (annotation: of exponential growth; then) also causal, and analytical, then: ${\widehat {f} }(i\tau )=F(-2\pi \tau )$. Thus, extending the Fourier transform to the complex domain means it includes the Laplace transform as a special case in the case of causal functions—but with the change of variable _s_ = _i_<-- markdown separator -->2π<-- markdown separator -->_ξ_. (annotation: That is, make any expression containing _ξ_ in the form of that on the right hand side, then simply replace it with _s_. Now the Fourier transform with _ξ_ as input becomes the Laplace transform with _s_ as input.)
Output: If _f_ is {@{(annotation: of exponential growth; then) also causal, and analytical}@}, then: {@{${\widehat {f} }(i\tau )=F(-2\pi \tau )$}@}. Thus, extending {@{the Fourier transform to the complex domain}@} means it includes {@{the Laplace transform as a special case}@} in the case of {@{causal functions}@}—but with {@{the change of variable _s_ = _i_<-- markdown separator -->2π<-- markdown separator -->_ξ_}@}. (annotation: That is, make {@{any expression containing _ξ_}@} in the form of that on {@{the right hand side}@}, then simply {@{replace it with _s_}@}. Now {@{the Fourier transform with _ξ_ as input}@} becomes {@{the Laplace transform with _s_ as input}@}.)

Input: From another, perhaps more classical viewpoint, the Laplace transform by its form involves an additional exponential regulating term which lets it converge outside of the imaginary line where the Fourier transform is defined. As such it can converge for at most exponentially divergent series and integrals, whereas the original Fourier decomposition cannot, enabling analysis of systems with divergent or critical elements. Two particular examples from linear signal processing are the construction of allpass filter networks from critical comb and mitigating filters via exact pole-zero cancellation on the unit circle. Such designs are common in audio processing, where highly nonlinear phase response is sought for, as in reverb.
Output: From another, perhaps {@{more classical viewpoint}@}, {@{the Laplace transform by its form}@} involves {@{an additional exponential regulating term}@} which lets it converge {@{outside of the imaginary line}@} where {@{the Fourier transform is defined}@}. As such it can {@{converge for at most exponentially divergent series and integrals}@}, whereas {@{the original Fourier decomposition cannot}@}, enabling {@{analysis of systems with divergent or critical elements}@}. Two particular examples from {@{linear signal processing}@} are the construction of {@{allpass filter networks from critical comb}@} and {@{mitigating filters}@} via {@{exact pole-zero cancellation on the unit circle}@}. Such designs are {@{common in audio processing}@}, where {@{highly nonlinear phase response}@} is {@{sought for, as in reverb}@}.

The remaining examples follow the same pattern; open the files for more entries if needed.

Heuristics and rules

The form of the examples above is complemented by a set of practical heuristics, which are now incorporated below under the appropriate sections (learned rules no longer require a separate file).

General editing principles

• Preserve the source verbatim except for cloze markup; do not paraphrase or reflow text.
• Equations stay whole: wrap an entire $...$ or $$...$$ block in a single cloze; never split math.
• Prompt length is irrelevant for calculations. When constructing two‑sided or one‑sided cards involving numeric problems, copy every number, parameter, and even full equations to the left of the separator; the prompt may be arbitrarily long. The warning rules in the validator are designed to catch cases where the answer cannot possibly be deduced from the prompt, not to force brevity. Resist any urge to split a single calculation across multiple cards purely to shorten the left-hand text – that reduces clarity and defeats the purpose of spaced repetition.
• Anchor context: leave visible words around each deletion to give a hint; avoid blanking a sentence entirely unless context is crystal clear.
• Mirror the user’s style when they supply examples or corrections.
• Preserving HTML entities and escapes exactly; treat them as opaque literals.
• Submodule safety: never edit private/, self/, or other submodules without explicit authorization.

Sentence and clause handling

• Work paragraph by paragraph; add clozes only within a single paragraph before moving on.
• Break long sentences with multiple ideas into separate cards rather than one huge deletion. Semicolon lists get individual clozes per clause.
• Split around contrastive conjunctions (but, however, etc.) and keep conditional connectors (if, when, etc.) visible unless the condition itself is tested.
• Default to inline clozes; use ::@:: or :@: only if the user requests QA style or an example clearly shows it.
• When expanding an already-clozed official solution block, do not just add more equation clozes. Prefer a balanced trio of recall targets: what is being counted/conditioned on, why the method applies, and what the final conclusion is. This is especially useful in academic question banks where the formulas are short but the decisive explanatory phrase carries the real memory cue.
• For simple declarative sentences, consider hiding subject and object separately to yield focused cards.
• Articles, possessives, prepositions, and qualifiers should be included inside the cloze when they are part of the tested concept.
• Hide the minimal meaningful semantic unit—adjust if the user later shifts words in or out of the cloze.
• Only cloze a year or date if it is central to the idea; incidental dates remain visible.

Lists and section‑wide edits

• Two‑sided QA lists: when the entire section should become a list of question/answer pairs, prepend a horizontal rule and the exact phrase “Flashcards for this section are as follows:” on a blank line, then convert each bullet separately.
• Treat adjacent paragraphs as a unit if the user edits them together; record such multi‑paragraph patterns as new examples.

Improvement process

• After editing, solicit feedback (“which deletions are wrong?”, “prefer QA instead?”) and use corrections to update examples and rules immediately.
• Analyse user edits with git --no-pager diff --word-diff --no-color and add training examples for boundary shifts or merged clozes.
• Summarise learned rules periodically into this document; avoid letting session memory grow unbounded.

Implementation notes

• The skill may be purely heuristic; there is no need to parse Markdown deeply or run pytextgen. Simple text manipulation is sufficient if the results are usually reasonable and users can refine them.
• Only edit visible Markdown notes; do not modify private/ or any submodule content without explicit permission.
• Dialogue should remain interactive and improvement-driven; remember past corrections to improve future performance.

Using this skill makes flashcard creation faster while faithfully following the user’s style. Always start by reading the example section—those patterns are the ultimate reference.