← ClaudeAtlas

studylisted

Daily study driver — reads due-card state and course progress, plans today's session (reviews → quiz → new material), runs it with the user, and writes the study log.
TimCodes04/second-brain · ★ 1 · AI & Automation · score 77
Install: claude install-skill TimCodes04/second-brain
# /study [--minutes <N>] The daily loop. Read `docs/pedagogy.md` (session rules) and `docs/vault-schema.md` first. Vault: `<vault>`. ## 1. Gather state (read-only) - **Due cards:** in every card file (definition, parsing recipe, and the local-day cutoff are at the bottom of `docs/vault-schema.md`), grep `<!--SR:` and count per deck: due today/overdue, plus cards with no SR comment (= new). - **Due problems:** grep `retry after` in `courses/*/problems/*.md`; collect problems with retry date ≤ today. - **Inbox:** count pending items in `Inbox.md`. - **Courses:** active course homes (`lessons_done`/`lessons_total`, `target_date`); lessons `status`/`mastery`/`last_quizzed`; quiz logs from the last 14 days; yesterday's study log carry-over. ## 2. Plan the session Budget: `--minutes` or ~45 by default. Order per pedagogy: ⓪ process pending Inbox items (follow the `/inbox` procedure) ① due-card review in Obsidian (~1 min per 4 cards) ② retrieval work in the terminal — due problem re-attempts (follow the `/problems --due` procedure) and/or `/quiz --weak` on the weakest/stalest lessons ③ new material (next unread lesson, `/learn --continue`, or next `/ingest` chunk). Near a `target_date` (< 3 weeks), shift ③ toward cumulative interleaved quizzing instead of new content. Print the plan, e.g.: > ① Obsidian: review 23 due cards (`qm-sakurai` 17, `lagrangian` 6) — do this now, then come back. > ② `/quiz qm-sakurai --weak` (L03 mastery 45, stale since Aug 12) — I'll run it here. > ③