| name | exam-cheatsheet |
| description | 全员通关后把 错题本+笔记本+知识点窗口+wiki 编译成考前速记小抄 cheatsheet.md（每条要点带可溯源 锚点），并在视觉产物模式或用户明确要求 PDF/打印版时按指定页数渲染成打印级 PDF：按「必背结论/公式 → 有难度例题（必要时含题面图）→ 例题解答（代入公式、保留基础过程）→ 要点解释（同类题怎么办）」 四段组织。当复习收尾、用户要「考前小抄/速记/总结/打印版」时使用。
|
| license | MIT |
exam-cheatsheet — pre-exam cheatsheet compiler
Purpose
Compile, rather than free-generate, mastered content into workspace-root cheatsheet.md. Every top-level bullet must link into notebook/, mistakes/, or references/wiki/. Do not teach new material or invent questions. Render the requested-page-count PDF only for standing visual mode or an explicit PDF/print request. Never write the retired walkthrough.md; leave an existing copy untouched.
Activation
Trigger on an explicit request for 「考前小抄 / 速记 / 总复习」, or when review is wrapping up after all phases and persisted artifact_mode=visual. Automatic final review under chat stays a conversational exam-review summary.
Inputs
- Weak-spot source:
study_state.json (mistake_archive, confusion_log, and phase_checklist) when it exists; otherwise the possibly stale generated study_progress.md. Read these first, then mistakes/index.md and notebook/index.md when present; their full entries provide preferred ready-made anchors.
- Rank
knowledge_window status out_window above in_window and verified (codes are defined by scripts/i18n.py).
- Read core conclusions and formulas from every mastered chapter in
references/wiki/, derived from study_state.json's current_phase/phase_checklist when it exists, otherwise study_progress.md, checked against study_plan.md. Lazy-load one chapter at a time.
- Use
references/quiz_bank.json for teacher-flagged items and answer frameworks. Resolve scripts/select_hard_questions.py from ${CLAUDE_SKILL_DIR}, never the student workspace; it returns a flat ranked list which the agent groups by knowledge point.
Workflow
- Gate artifacts. Read
study_state.json.artifact_mode; missing, legacy, or unknown means chat. Never infer a subscription tier or add a fourth required first-contact question. Automatic chat review creates no sheet; an explicit sheet request may create Markdown. Only standing visual or an explicit one-shot PDF/print request authorizes rendering. A one-shot request does not modify the persisted value. Never install dependencies or skills silently.
- Build the skeleton. Weak spots come first. Per chapter retain only high-frequency or high-scoring formulas, conclusions, and one-sentence definitions.
- Select one hard example per key point. For each mastered chapter run
python "${CLAUDE_SKILL_DIR}/scripts/select_hard_questions.py" --workspace <ws> --chapter <N> --mode 查缺补漏 -n <M> --json. Both --chapter and --mode are required: they avoid a missing-range failure in 某章起步补弱 and override easy-first 零基础从头讲. Set <M> at least to the bank length so the default top ten cannot starve later points. Group the flat result, prioritize points linked to mistakes/confusions, and choose the hardest candidate per point. With no linked bank item, emit 「无题库例题」 and only the 「必背结论/公式」 and 「要点解释」 sections; never invent a replacement.
- Fail closed on prompt assets. For
requires_assets=true or maybe_requires_assets=true, embed every question_context, figure, diagram, and table as workspace-relative references/assets/ links, labeled 题面图 for zh/bilingual or Question-side asset for en. Preserve but never embed student_attempt; one declaration taints the same physical path across the complete quiz, teaching, and content-unit layers, including a duplicate official-looking declaration. Missing or unusable assets require a self-contained alternative. A stub or page_reference item likewise needs its original-page render or replacement by a full item. Never include an example whose prompt figure/page is invisible. cheatsheet_render.py performs the shared three-layer policy and canonical-path gate; do not bypass it with a custom Markdown/image renderer.
- The worked solution states the formula, substituted values, and result; only intermediate arithmetic may be omitted. The takeaway starts with the recognition cue and then the answer framework. Material-backed lines may remain unlabeled; AI supplements require 🟡 AI补充，可能与你老师讲的不完全一致, AI answers require ⚠️ AI生成答案，非老师/教材提供, and missing/unknown bank answer provenance requires 「来源未知」. Do not let uncertain content inherit the material default; see .
Output Contract
cheatsheet.md uses active-language headings per mastered chapter: zh uses 「必背结论/公式」→「例题」→「例题解答」→「要点解释」; en uses Must-memorize conclusions & formulas → Worked example → Worked solution → Takeaway. Every bullet is traced and validation passes.- An explicit
chat request delivers Markdown only unless it also requests print/PDF. Authorized rendering delivers exact-page-count cheatsheet.pdf, or cheatsheet.html plus print instructions on the no-browser path.
- Student-facing output defaults to English (Simplified Chinese if the student opened in Chinese). Persisted
language values zh, en, and bilingual select single-language or mirrored output per docs/language-policy.md.
Language packs
Display aliases are normalized to zh, en, or bilingual.
Boundaries
- Unsupported content needs the applicable 🟡 or ⚠️ label. The sheet compresses completed review; it never bypasses source labels or the
quiz_bank-only rule.
