flashcard-v3.6 - Improved anki-flashcard-generator skill

anki-flashcard-generator v3.6

Major rebuild of the skill around a single design principle: rules block defects, examples convey style. Testing across model versions showed that prescriptive style rules (rigid atomicity targets, mandatory "why/how" framing, universal bidirectional cards) globally distorted deck character, while defect-blocking rules cost nothing. v3.6 restores the terse, exam-aligned register of the original prompt versions while keeping the two genuine gains of v3.3–v3.5: accuracy verification and targeted interference management. The skill is roughly one third the length of v3.5.

Changed

• Atomicity softened to "one idea per card, judged flexibly" — a definition may bundle one directly associated detail (its formula, unit, or key property) when naturally recalled together. Never a reasoning chain or second independent concept
• New bundle-or-split rule — a detail bundled into a definition must not also get its own card, and no answer may contain another card's answer. Closes the pay-twice duplication observed in earlier versions (e.g. "measured in Pa" inside a definition alongside a separate unit card)
• Reverse cards scoped to key terms only (terms the exam asks candidates to define), replacing the universal bidirectional rule that doubled deck size. Scoping self-adjusts by deck type: sparse on topic notes, near-complete on definition sheets
• Interference check constrained to exactly two cases — contradicting cards (rephrase to name context) and pairs the source itself contrasts (one compare card stating the point of divergence). Compare/contrast is no longer a free card type; duplicate compare cards are banned
• Verification step reworked — clear factual errors are fixed; possible syllabus-level simplifications are kept and flagged in the chat response, never silently rewritten and never written into the output file
• Explain cards demoted — only where the source itself explains the reasoning and it is a likely exam point, mechanism stated in at most two sentences. Recall content is never converted into explain cards
• Card-count anchor removed — deck size is determined by content density; the material decides

Added

• Pipe-character exclusion — | is banned inside questions and answers (write "magnitude of v", not |v|), preventing silent import corruption
• Circular-answer exclusion — answers that merely restate the question are omitted and flagged (a defect that survived four prompt/skill generations before being named)
• Brevity tiebreaker — when any style rule conflicts with brevity, brevity wins. Resolves the rule conflict that drove v3.5's answer padding
• Bad→good rephrase example on the yes/no ban — the one rule observed to slip under v3.5-era instructions now teaches by example
• Bundled-definition example (impulse: definition + Ft = Δp) and a non-physics example (binary search complexity) signalling domain generality

Removed

• Depth of processing, personal connection, dual coding, and understand-first principles, plus all retention-science rationale prose — identified as the primary drivers of register drift
• The 6-second answerability and 25-word answer targets
• The quality checklist (redundant with the rules)
• All GCSE-era "Higher Tier" references

Validation

Tested on AQA A-Level Physics sources (Classification of Particles topic notes; Mechanics & Materials definition sheet), run both inside and outside Claude projects:

• Outputs converged to near-identical decks across project/no-project conditions, demonstrating the skill dominates both model defaults and project memory
• Controlled comparison against a bare minimal prompt on the same model showed the skill suppresses model-default elaboration: unsolicited verification/worked cards, yes/no questions, and bundle-plus-duplicate redundancy all appeared without the skill and not with it
• Every defect class identified across v2–v3.5 decks (answer padding, answer leakage, deck doubling, duplicate compare cards, unit-leak duplication, circular answers) is fixed or reduced to single-line generation variance

Companion artifacts

• anki-flashcard-prompt-v5 — standalone fallback prompt with full instruction parity, for use when the skill feature is unavailable. Keep in sync with SKILL.md
• README updated to match v3.6, including the defect-blocking vs style-by-example design rationale and per-version notes

Upgrading

Drop-in replacement for the existing SKILL.md — same name, same trigger phrases ("Anki flashcard" / "Anki deck"), same output format. Decks generated under v3.3–v3.5 remain import-compatible; expect new decks to be terser, smaller, and free of the padding and duplication patterns above. If project memory contains card-design principles from earlier versions, remove them and defer to the skill as the single source of truth.