Spaced Review

Studying something once and marking it mastered in TRAINING.md doesn't mean it stays mastered — this skill resurfaces material on a schedule so it actually sticks, using a simple Leitner box system rather than one-shot quizzes.

Anki vs spaced-review

Two spaced-repetition systems run in this project on purpose, split by what they're each good at — don't let them duplicate each other:

• Anki (via quiz-flashcards' CSV export) owns atomic recall: terms, definitions, API parameters, short facts. Anki's own scheduler decides when those are due — this project has no visibility into Anki's review state, so never try to reproduce or second-guess it here.
• spaced-review/REVIEW.md owns everything Anki's card format can't grade: open-ended "explain this to me," teach-it-back, and system-design-style recall where the answer isn't a short fixed string. Only put items here that genuinely need that format — don't add a REVIEW.md row for something that's really just a fact better suited to a flashcard.

If the user says a topic is "in Anki," treat that as their fact-recall coverage being handled — don't nag about it from this skill, and don't create a REVIEW.md row duplicating flashcard-style questions for it.

The schedule (REVIEW.md)

REVIEW.md at the project root is the source of truth — one row per reviewable item (a train-topic topic, a book-study chapter, or anything else validated). Columns: item (matches the topics/<slug>/ or books/<slug>/chNN reference), box (1-5), next_due (date), last_result (pass/fail), streak.

Box → interval: Box 1 = 1 day, Box 2 = 3 days, Box 3 = 7 days, Box 4 = 16 days, Box 5 = 35 days. A correct, confident recall moves the item up one box (longer interval); a wrong or shaky answer resets it to Box 1 regardless of where it was.

REVIEW.md doesn't create its own entries — train-topic and book-study add a row (starting at Box 1) the first time an item is validated (quiz passed or exercise completed). If you're running this skill and an item that should be tracked isn't in REVIEW.md, add it starting at Box 1 rather than skipping it.

Running a review session

1. Read REVIEW.md, filter to rows where next_due <= today.
2. If nothing is due, say so plainly and stop — don't manufacture a review session out of material that isn't due yet. Optionally mention what's coming up soonest.
3. For each due item, ask 2-4 closed-book, active-recall questions (short answer or teach-it-back, not multiple choice — recognition is easier than recall and defeats the point). No peeking at topics/<slug>/README.md before answering.
4. Ask for a quick confidence rating (low/medium/high) alongside the answer — mismatched "high confidence, wrong answer" is worth calling out explicitly, since that's the failure mode spaced repetition is best at catching.
5. Grade honestly. Partial/vague recall counts as a miss for scheduling purposes even if it's "basically right" — the point is testing effortless retrieval, not partial credit.

After the session

• Update each reviewed item's row in REVIEW.md: new box, new next_due (today + interval), last_result, and streak (consecutive passes; reset to 0 on a miss).
• If an item fails repeatedly (2+ resets), flag it back to the user and suggest a re-run of train-topic's deep-dive/hands-on stages rather than keep drilling it in isolation — spaced review surfaces forgetting, it doesn't fix a genuine gap in understanding.
• Don't touch TRAINING.md's mastered status based on a single miss here — that's still train-topic/book-study's call, this skill only manages the review schedule.