vocabulary-builder

name	vocabulary-builder
description	Use when the user wants to learn, review, or be quizzed on English vocabulary for exams: leveled word sets (B1-C2), collocations, academic vocabulary, phrasal verbs, or spaced-repetition review of previously studied items ("quiz me on my words", "new C1 vocabulary", "review my vocab"). Maintains a Leitner-box review state on disk and logs quiz results to the progress log.

Vocabulary Builder

Leveled vocabulary with collocations and a file-based Leitner system. Paths are relative to ${CLAUDE_PLUGIN_ROOT} (if unset, resolve relative to this file). The review state lives in the user's progress directory ($EXAM_COACH_HOME or ~/english-exam-coach/), NOT in the plugin.

When to use

The user asks for new vocabulary at a level or topic, asks to review/quiz existing vocabulary, or meets unknown words in another drill and wants them kept.

Steps

1. New set: identify level (B1–C2) and theme (topic, exam register, e.g. "hedging for C1 writing"). Follow the structure in data/item-bank/seed/vocabulary-sets.md: 5–8 items per set, each with collocations and one original example sentence. Level the words, not just the theme: pick items whose frequency/register fits the level — B1 high-frequency everyday; B2 common academic (roughly the AWL mid-bands); C1 lower-frequency and abstract; C2 idiomatic/precise nuance — and write the example sentence at that level too, so the set carries level-appropriate language beyond the headword. Note that hedging and precision-verb sets sit at C1/C2, not B1–B2. Open lists like the Academic Word List may be referenced nominatively.

2. Persist for review: append new items to <base>/vocab-box.json (create if missing). The file is a JSON array; each element is {"item": ..., "level": ..., "box": 1, "added": "YYYY-MM-DD", "due": "YYYY-MM-DD"} with ISO dates. New items enter box 1 with due = today, so they surface in the next review round. Canonical Leitner intervals (must match vocabulary-sets.md): box 1 = next session · box 2 = 1 day · box 3 = 3 days · box 4 = 7 days · box 5 = 30 days (mature). This file is working state and MAY be rewritten — unlike attempts.jsonl, which this skill never touches directly.

3. Review ("quiz me"): read vocab-box.json, select items with due <= today (oldest first, max ~12 per round). Quiz actively — ask for the word from a definition/gap sentence, or a collocation completion; a plain "do you remember this?" is not a test. Correct → box +1 (max 5), set due = today + the new box's interval above; wrong → back to box 1, due = today (next session). Rewrite the file once at the end of the round.

4. Score the round (n correct / n asked) and show which items fell back to box 1.

5. Log the round (silently):

   python3 "${CLAUDE_PLUGIN_ROOT}/skills/progress-tracker/scripts/log_attempt.py" \
     --exam <target-exam or cefr level file, e.g. cefr-c1> \
     --skill vocabulary-builder --task-type spaced-review \
     --level <level> --score <n> --max <asked> --seconds <time>
   

   (On Windows without python3, use python or py instead.) When an item is missed for a reason worth re-testing beyond the word itself — a wrong form, a collocation, a register slip — log it with log_error.py under vocabulary and run queue.py sync; the word stays in the Leitner box, the pattern enters the queue. See ../progress-tracker/references/the-loop.md. A due-based round often mixes levels, but the log takes one --level and one --exam: set --level to the modal (most common) level of the items quizzed and --exam to the user's target exam (or the cefr-<level> file matching that modal level) — for a mixed round the review level is only a rough anchor. For a new-set study session without a quiz, do not log — only tested recall counts as an attempt.

Boundaries

• Word lists and example sentences are original or from open sources named nominatively; never proprietary coursebook or exam-board vocabulary banks.
• vocab-box.json is the only file this skill rewrites; never touch attempts.jsonl except through log_attempt.py.
• Do not log rounds that didn't happen; unanswered items count as wrong only if the user gave up, not if the session was cut short.