Exam Quiz Agent Skill

exam-quiz — question drilling and grading

Purpose

Present one chapter/phase-scoped bank item at a time, grade against its stored answer, archive wrong/skipped items through state, and return control to exam-cram. Never invent a question or answer.

Activation

Use after teaching when a checkpoint is needed, or when the student asks for drills or a mock exam.

Inputs

• Existing references/quiz_bank.json, whose items have type, answer/provenance fields, and chapter or phase; subjective items also have keywords.
• Current chapter/phase and study_state.json mastery/scope. An untagged item cannot enter a chapter checkpoint.
• Optional difficulty (1–5) and difficulty_reason from score_difficulty.py: a structural lower bound, never semantic truth or a per-student score.

Workflow

• Select only eligible bank items. Filter both chapter and phase. A missing bank is an incomplete workspace and returns to exam-ingest; an existing but empty usable pool produces no substitute and caps completion at covered_unverified.

The default source pool is mixed. Persist a student restriction and select it with scripts/select_questions.py; exclude and count items lacking source_type. Before any one-turn exception say 「⚠️ 临时覆盖你的 <scope> 范围偏好」 or ⚠️ Temporarily overriding your <scope> scope preference; do not silently change the stored scope.

For targeted/checkpoint selection run python "${CLAUDE_SKILL_DIR}/scripts/select_hard_questions.py" --workspace <ws> --chapter <current> -n <k>. --chapter is the only exact chapter filter; --from-chapter N means every numeric chapter ≥N and is only for shore_up, never a checkpoint. Explicit cross-chapter practice may omit chapter. The selector combines structural difficulty (using score_difficulty.py on the fly when needed) with mistake/confusion/window mastery, mode, and stored scope. fill_gaps serves weak points 先易后难, then mastered items 先难挑战; from_scratch is globally 先易后难. shore_up requires explicit chapter/from-chapter. Ordering is deterministic, not LLM ranking.

• Show prompt assets first (fail-closed). For requires_assets=true or maybe_requires_assets=true, before asking, explaining, hinting, or solving, actually render every question-side question_context / figure / diagram / table asset, labelled 题面图 or Question-side asset. A path is not an image. Show answer_context / worked_solution only later, labelled 答案图 or Answer-side asset. Preserve but never display student_attempt: one occurrence taints the same physical path across the complete quiz, teaching, and content-unit layers, so an official-looking duplicate declaration is also unusable. Missing/unreadable files block the structured workspace; an existing asset that the UI cannot render causes an item-level skip. Prefer a safe, self-contained full item. stub and page_reference also require the prompt asset or original page first. Always use python <package-root>/scripts/show_question_assets.py --workspace <ws> --id <qid> --lang <zh|en> so the shared three-layer policy is applied; exit 1 means skip. Do not bypass it by rendering a raw bank path yourself. See docs/file-format.md §4.

• Grade by type. choice: stored option. subjective: required keywords/steps with equivalent wording accepted and coverage reported. fill_blank: stored fill with valid synonyms. true_false: verdict plus one-line reason. code: required edits/output. diagram: run the standard algorithm from render_hint, derive the structure, then compare; teacher convention prevails.

• Use the escape hatch. First wrong answer gets the logic gap, stored explanation, and a hint. On the second consecutive wrong answer offer view hint / skip and archive / continue.

• Persist evidence and feedback. Before any write, if study_state.json is absent and Python works, run python "${CLAUDE_SKILL_DIR}/scripts/update_progress.py" --workspace <ws> init; only when Python truly cannot run may the generated Markdown be maintained directly. For every handled item record record-phase-evidence --kind checkpoint --ref <qid> --outcome passed|wrong|skipped; an ID alone is not mastery. Wrong/skipped items also use python "${CLAUDE_SKILL_DIR}/scripts/update_progress.py" --workspace <ws> add-mistake --id <qid> --chapter <ch> --note <reason>. A nonzero state command is a fail-loud write error, not permission to edit the generated view.

Before replying, pipe full verdict, gap, explanation, and source line to python "${CLAUDE_SKILL_DIR}/scripts/notebook.py" --workspace <ws> add-entry --chapter <ch> --type feedback --id <qid> --title <gist>. Same chapter/id replaces in place. Wrong/skipped feedback also passes --mistake to mirror mistakes/chNN.md; that supplements, never replaces, the state row. Then send a short digest and language-pack link. If notebook writing fails, say so and give the full feedback in chat; file-less clients use chat/text breakpoints.

• End every graded item with one source line: 题目来源：<file/page/source_type>｜答案来源：<material/AI>｜<label> or Question source: <...> | Answer source: <...> | <label>. Missing metadata says 「来源未知」 / Source unknown (or Source page unknown), never an invented filename/page. The label is one complete canonical sentence from docs/language-policy.md: 🟢 来自资料; 🟡 AI补充，可能与你老师讲的不完全一致; or ⚠️ AI生成答案，非老师/教材提供, with its English counterpart. When no material answer exists, both the 解析/参考答案 title and source line carry the full ⚠️ sentence; without a stored answer, do not force a verdict.

Output Contract

• One item at a time; pass/not-pass plus key-point feedback; finish with the source line and refreshed progress panel.
• Persist feedback before the digest; wrong/skipped items need checkpoint evidence, state mistake row, and notebook mistake mirror.
• exam-cram / exam-tutor, not this skill, calls evidence-gated complete-phase.
• Student prose follows the persisted language with single-language purity: English by default, Simplified Chinese if the opening was Chinese, or explicit bilingual blocks.

Language packs

Load before student-visible output:

• 中文 → ../../locales/zh/skills/exam-quiz.md
• English → ../../locales/en/skills/exam-quiz.md
• 双语 → compose both blockwise, zh then > EN:, under docs/language-policy.md

Display aliases are normalized to zh, en, or bilingual; unset language follows the merged first ask.

Boundaries

• study_state.json is the source of truth. Update it only via python "${CLAUDE_SKILL_DIR}/scripts/update_progress.py" --workspace <ws> ...; study_progress.md is generated. Fail writes loudly; initialize state whenever Python works.
• Never create a replacement item, invent a source/answer, grade a diagram from memory, or serve a visual-dependent prompt whose image was not shown.
• For visual statistics, report both quiz-bank visual items via scripts/list_image_questions.py (total/requires/maybe/suspects) and material figure pages via scripts/list_figure_pages.py. If the index is absent, build it with scripts/build_visual_index.py; never count by hand.