Imported from SousekiL/astam-review-notes (SKILL.md). Install upstream with npx skills add SousekiL/astam-review-notes. Copyright stays with the author.

ASTAM Review Notes Skill

When this applies

Use this skill when the user asks for ASTAM / STAM-style summary, problems, solutions, or mistake content for SOA exam preparation.

Step 0 — Load context from the knowledge base

Before generating or editing review content, consult reference materials in this skill. For Type = problems, follow this priority order:

Syllabus topic codes (for filenames and progress)

Map every request to one primary topic T1–T6 (SOA ASTAM syllabus areas and approximate weights):

If the user’s input fits multiple topics, pick the dominant one for the filename; mention secondary tags in the note body (YAML frontmatter or a short “Also relates to” line).

Practice analytics — dashboard + JSON ledger (English dashboard)

Maintain a simple numeric ledger and an English rollup note so progress is visible without a separate web app:

When to update (agent — required):

1. Type = problems: Append problemSets (id, recordedAt, topic, title, format, vaultNote, pointsPossible, optional subparts / largeQuestions, optional subBranchIds e.g. ["t2-a","t2-b"]). Bump lastUpdated; refresh At a glance and Problem sets.
2. Type = mistakes / grading: Append gradedSessions (plus optional subBranchBreakdown, subBranchScoringMode, questionBranchMap). Sub-branches may overlap: if one large question tests multiple outcomes, credit its full earned/possible to each listed branch (non_exclusive_by_question); then branch possible sums can exceed the session’s pointsPossible — state this in the dashboard. Recompute topicTotals (session points, exclusive) and subBranchTotals from subBranchBreakdown. Refresh the dashboard tables. Legacy sessions with empty subBranchBreakdown only affect topic rollup.
3. Syllabus change: If SOA rewrites outcomes, edit syllabusBranches in the JSON and realign dashboard labels from 2026-spring-exam-astam-syllabus.md (or the current syllabus file).

Keep JSON and dashboard consistent in the same edit batch. Do not leave the dashboard stale after changing the ledger.

Atomic update checklist (agent — same task, same save):

1. Edit ASTAM-Practice-Data.json (problemSets and/or gradedSessions, totals).
2. Immediately update ASTAM-Practice-Dashboard.md tables / “At a glance” so every number shown has a JSON source.
3. If you only change one file by mistake, re-open the pair and fix before ending the turn.

Optional later: a small script could render dashboard rows from JSON; until then, the checklist prevents drift.

Syllabus subpart mapping (authoring → Mistakes → ledger)

Goal: Every scorable subpart should map to t#-a / t#-b / … (ids in ASTAM-Practice-Data.json → syllabusBranches) so that 错题 / 得分 can be pasted into gradedSessions without re-reading the whole stem.

Rules for the agent when Type = problems:

1. Do not put syllabus codes in the visible stem (no “(T3b)” in the text the student sees).
2. Do record the mapping in a machine-friendly, low-noise way. Preferred (Obsidian preview hides it): an HTML comment immediately after each subpart heading or prompt line, e.g. <!-- astam: t3-b -->. Alternative: a single block at the end of the Problems note, e.g. <!-- astam-syllabus-map: {"Q1":{"(a)":"t3-a","(b)":"t3-b"}} --> (JSON inside the comment).
   • Anti-forget template: When drafting a new sheet, insert a placeholder immediately after every scorable subpart prompt, e.g. <!-- astam: t?-? -->, then replace ? once ASTAM-Practice-Data.json branch ids are chosen. Optional vault habit: an Obsidian template/snippet that pastes <!-- astam: --> so the cursor lands between the colons.
3. When the user later asks for Mistakes or you append gradedSessions, copy branchId from that map into mistakeSubparts (and/or questionBranchMap for formal MCQ), then recompute topicTotals and subBranchTotals (or subBranchBreakdown per session). Overlap: if one subpart truly tests multiple outcomes, use the same non-exclusive pattern as T2 MCQ (subBranchScoringMode, multiple branches).

gradedSessions shape (for humans + agents): See ASTAM-Practice-Data.json (in user's vault) keys _gradedSessionShapeGuide (field glossary) for structure examples.

When Type = mistakes: Prefer listing each wrong/graded subpart with ref + branchId in JSON so the dashboard branch rows stay reproducible.

Agent responsibilities (on Dashboard study-plan section):

1. Whenever you create or rename a syllabus-related note in the user's vault, update the Dashboard's T1–T6 index and weekly rows with [[wikilinks]].
2. If the user reports review progress, set the matching 完成 checkboxes to [x] in the Dashboard weekly tables.
3. Bump the 最后更新 line at the end of the study-plan section.

Note: When a topic has several practice sheets in syllabus / study order, use a numeric segment after T# so the file explorer sorts them: T5-1-bayesian-credibility-Problems.md, T5-2-buhlmann-straub-Problems.md, T5-3-empirical-bayes-Problems.md (same pattern for T2–T4 multi-sheet sets).

File naming (syllabus-friendly, sortable)

Prefer ASCII-friendly names so folders sort by topic then subtype:

T{n}[-{k}]-{ShortTopicSlug}-{Kind}.md

• n: 1–6 per table above.
• k (optional): Review-order index when multiple sheets cover the same T# in syllabus sequence (e.g. 1 = first outcome block, 2 = next). Example: T4-1-parametric-mle-ci-Problems.md, T4-2-delta-bayesian-Problems.md, T4-3-model-selection-gof-Problems.md.
• ShortTopicSlug: short English kebab-case hint (e.g. severity, aggregate-recursion, coverage-deductible, parametric-selection, buhlmann, pricing-reserving). For subsections, extend the slug (e.g. aggregate-ii-recursion).
• Kind: one of Summary, Problems, Solutions, Mistakes.

Examples: T2-1-aggregate-iii-Problems-MCQ.md, T4-2-delta-bayesian-Problems.md, T5-3-empirical-bayes-Problems.md, T5-2-buhlmann-Summary.md.

Problems naming (MCQ vs WA): Use the same …-Problems.md for both default (checkbox MCQ) and 非选择题 (WA). Do not add a -MCQ segment to the filename (e.g. not …-Problems-MCQ.md). If you need two sheets for the same topic (one MCQ, one WA), distinguish with ShortTopicSlug (e.g. aggregate-iii vs aggregate-iii-wa). Optional separate answer file: …-Problems-Key.md; mistake log: …-Problems-Mistakes.md.

Legacy files with spaces, …-Problems-MCQ.md, or older T2 Aggregate II … names may remain; for new notes, use this pattern. When creating or heavily editing a legacy file, add an HTML comment at the top (after YAML if any), e.g. <!-- canonical: T2-aggregate-iii-Problems.md (MCQ) -->, so agents know the preferred rename target without breaking existing [[wikilinks]] until the user renames.

Each note should include near the top (after optional YAML):

• ASTAM topic: T# — [English syllabus name]
• Reference to relevant sources (ASM lessons, Past Exams, etc.) when applicable.

Answer key (…-Problems-Key.md) — isolation + when to create

Goal: The student can practice without seeing letters or worked MCQ answers; grading has a stable reference so the agent does not re-derive everything from scratch.

Placement (pick one — user may state a preference):

1. Preferred — separate file: T#-…-Problems-Key.md sits beside T#-…-Problems.md. The Problems note links it only in header/footer, e.g. After attempt: [[…-Problems-Key]]. No answer letters appear in the Problems file.
2. Single-note option (e.g. “Key 都放在最后”): If the user asks for one file, structure strictly as:
   • Top → middle: only stems, options (MCQ), study-guide highlights, syllabus-map comment — zero spoilers.
   • Bottom: a single terminal block separated by --- and a visible spoiler heading, e.g. ## Answer key (spoilers — after attempt only), optionally preceded by <!-- astam:stop-reading-for-practice -->. All correct letters, short rationale, and post-answer Knowledge highlights go below that heading only — never interleaved after each question in the scroll path used during timed practice.

Workflow norms (lowers “无 Key 批改成本”):

• After a Type = problems delivery, the student can say “生成 Key” in the same chat or the next message; opening the Key file is after attempt.
• On Type = mistakes / grading, prefer an existing …-Problems-Key.md when present; only re-solve from stems if Key is missing.

User parameters (mini protocol)

When the user omits parameters, infer reasonable defaults and state them in a one-line preamble.

Vault user norm (when Type=problems): Default to 选择题 — each scorable subpart is single-best-answer with exactly five options A–E (- [ ] **A.** … - [ ] **E.**). Use 非选择题 (constructed response, no A–E) only when the user explicitly requests it (e.g. 非选择题, 大题, written answer, WA, format: wa, constructed-response). If the user only says 出题 / practice / MCQ is unspecified, treat as MCQ.

Default when the user says 出题 / practice problems (Type=problems):

• Emit only the stem markdown: T#-{slug}-Problems.md (see naming above). Default answers: no: the Problems file contains no worked solutions and no correct letters (MCQ). Do not create …-Problems-Key.md unless the user explicitly asks (e.g. 答案、Key、详解、对答案、批改用 key).
• Format: unless the explicit 非选择题 / wa request applies, each subpart is single-best-answer MCQ with exactly five options A–E and Obsidian - [ ] / - [x] checkboxes.
• Exception — 非选择题 (explicit only): omit options; strict ASTAM constructed-response style with exam-like multi-level structure (see Content rules).

Content rules (short)

• Math (Obsidian-friendly — default for all new/edited ASTAM notes):

  • Inline: In stems, solutions, lists, and tables, use $...$ for every short formula or symbol. Never use $$ inside running text, bullets, or cells.
  • Display: Use $$...$$ only as a standalone paragraph (blank lines before and after).
  • Multi-line in one block: wrap in $$\begin{aligned}...\end{aligned}$$ (or gathered / cases as needed); use & at the start of each line or &= so lines align like printed notes — do not rely on raw \\ alone.
  • Preview: block left-aligned (optional): centered display is MathJax default; for flush-left blocks enable snippet astam-math-display-left.css in your Obsidian vault CSS snippets.
  • Avoid HTML entities inside math formulas.

• Type = summary: mainly Chinese; key terms 中文 + English.

• Type = problems — references & scope: Ground stems in ASM ASTAM + Past Exams (especially Past Exams for difficulty and complexity). Cross-check ASTAM/Syllabus/ so the set covers the syllabus objectives for the requested topic slice — no gaps in exam scope for that slice. Attribution: for each large question (or clearly per set if one theme), add a short line such as Adapted from: ASM Lesson n §… and/or ASTAM Season Year Problem k style (chapter / lesson / exam sitting as applicable). If purely synthetic, state Original practice item (syllabus: T# …).

• Type = problems — ASTAM exam structure: Mirror released ASTAM layout: typically one major question with (a)(b)(c)… subparts; subparts may include nested (i)(ii) given items. Adjust part count to match Past Exam density when possible; keep wording precise and mathematically check every numeric and formula before shipping.

• Type = problems — complexity calibration (mandatory):

  • Governing principle: Default difficulty is determined by Past-Exam-grade scenario depth + subpart dependency chains + proportion of comparison/judgment subparts — not merely by whether syllabus keywords are covered. ASM examples (not the chapter text) set the knowledge-combination and pitfall benchmark; Past Exams set the integrated-scenario and reasoning-length benchmark.
  • Scenario unity: Every large question must present one coherent scenario whose givens carry across subparts. Later subparts should explicitly depend on earlier results, data, or the same table/setup. Prohibited (unless user explicitly asks for "drill / 速刷"): packaging unrelated micro-calculations as (a)(b)(c)(d) under one question number.
  • Minimum reasoning depth (Medium floor): Each non-Easy large question must contain ≥2 subparts that require multi-step, non-formulaic reasoning — e.g. compare two distributions, select a model, interpret a hazard-rate crossing, evaluate an inequality, or apply a result from (a) inside (c). A subpart that is one direct formula plug counts toward the total but not toward this minimum.
  • Comparison / judgment quota: At least ~1/3 of all subparts across the set should be "Which / Determine whether / State with reasons / Select the preferred …" rather than pure "Calculate $X$".
  • ASM example alignment (mandatory search — not from memory): Before finalising any large question, consult ASM ASTAM Study Manual materials (if available in user's vault) and locate actual examples (by lesson number + example letter, e.g. "Lesson 2 Example 2H") whose knowledge combination or trap the question mirrors. Cite in the Attribution line: "Adapted from ASM Lesson X Example Y (…)". Prohibited: citing an ASM example purely from model memory without checking reference materials. Prioritize examples that integrate multiple concepts, not isolated definition checks.
  • Past Exam alignment (structure check): Before finalising, compare subpart count, given-block length, and dependency depth against one named Past Exam question of the same topic. If the draft is noticeably shallower (fewer givens, no cross-subpart dependency, shorter reasoning chain), add complexity until the gap closes. Cite the Past Exam sitting in the Attribution line.
  • Attribution for adapted / inspired content: When a question borrows structure, numbers, or a scenario twist from a specific ASM example or Past Exam question, the attribution line must state so — e.g. "Adapted from: ASM Lesson 2 Example 2H (mixture moments); ASTAM Fall 2023 Q2 (c)–(d) structure." Purely synthetic items state "Original (T# …)". Never copy verbatim; always change parameter values and/or scenario context.

• Type = problems (default — MCQ): English; 5–8 large questions unless count is set; each large question usually 3–5 subparts (flexible if mirroring a specific past-exam shape); each subpart = one MCQ with five options A–E (- [ ] **A.** … - [ ] **E.**). Stems: “You are given …”, Calculate / Determine / Which of the following. Difficulty: most items Medium+; include [Hard] on ~20–25% of large questions (see difficulty row). Never copy released exams verbatim — adapt with citation. Output: …-Problems.md only unless user requests Key. Do not emit …-Problems-Key.md unless asked.

  • Option typography parity (anti-spoiler, mandatory): Do not use bold/italics/underline/color/emoji to emphasize words in only one option, especially the correct one. In particular, never bold the correct answer's keyword text. Keep option-body typography visually consistent across A–E; only the option labels may use the fixed house style (**A.** … **E.**).
  • 🚩 Bookmark / flag per subpart — single canonical placement: After the subpart question text, on its own line, immediately before the first - [ ] **A.** option — never after option E and never before the question sentence. Exact sequence:

    (a) <!-- astam: t#-x --> *(2.5 pts)* [question text]
    
    - [ ] 🚩
    
    - [ ] **A.** …
    

    For WA (no A–E): place - [ ] 🚩 immediately after the prompt line, before any sub-items. No alternatives are permitted — always use this position so grading scripts can locate the flag unambiguously. The user checks - [x] 🚩 to bookmark for review even when answered correctly. When grading (Type = mistakes), flagged subparts are treated like wrong/blank for full inclusion; flagged-but-correct subparts appear in ## 2b and are not counted as errors in score tables.
  • Correct-option letter spread (MCQ): Across the whole problem set, avoid parking most correct answers on one letter (e.g. a long run of C). After assigning numeric/text keys, permute option rows so the correct A–E labels are roughly balanced over all scorable subparts (target no letter much above ~30% of subparts; for 18–24 subparts, aim for 3–5 corrects per letter when feasible). Add a one-line Authoring note in the Problems header if useful (“letters shuffled — do not guess by frequency”). Re-check that each subpart’s numeric/conceptual correct answer still matches the intended letter after reordering.
  • Default answers: no: no solutions, no correct letters in the Problems file. Knowledge highlights after each large question: Knowledge highlights (study guide — no answers) — do not reveal letters. If user later requests …-Problems-Key.md, add letters + brief rationale + highlights per reference-knowledge-highlights.md.
  • Conceptual / judgment MCQ — distractor design rules: For subparts that ask "Which / Determine whether / State which is preferred", the four wrong options must each represent a named, plausible misconception — not a vague opposite or obvious non-answer. Specifically: (a) at least 2 distractors should correspond to a common exam error documented in the T#-specific pitfall bank (reference-knowledge-highlights.md); (b) at least 1 distractor should be "correct in a different but related context" (e.g. the right formula for per-payment vs per-loss, or the right criterion for a different test); (c) avoid distractors that are eliminated purely by grammar or obvious illogic. After writing all 5 options, confirm each wrong option has a specific rationale for why a student might choose it.
  • MCQ numeric verification (mandatory before output) — Purpose: The correct option must be actually correct under a clear rounding convention — not a forced “closest to” choice because early rounding in intermediate steps pushed the true final value off every listed number.
    • Intermediate precision (multi-step numerics): Whenever a subpart chains two or more numeric operations, do not round intermediate results aggressively. Carry at least three decimal places at each step when writing intermediates (or keep exact fractions / full machine precision internally and only round the final answer to match how options are displayed). Too much early rounding is the usual reason the true final value no longer matches any option.
    • Authoring order: ① Derive the final numeric answer with the precision discipline above (full derivation in the Key / agent scratchpad). ② Then write the five options so one letter equals that final after the stem’s stated display rule (e.g. integer, nearest hundred, two decimals). ③ Build four distractors from named plausible slips (wrong formula, wrong tail factor, missing term, misread given). ④ Check that exactly one option matches the standard answer at that precision and that wrong options stay distinguishable without accidental ties.
    • “Closest to” stems: Still compute the high-precision result first. Prefer rephrasing or rescaling so the correct value appears exactly in the list when that fits the item. If the stem must say closest to, verify one option wins unambiguously from the high-precision result; if two options tie, change givens or option spacing before shipping.
    • If two options could both be defended as correct, revise the parameters before output.

• Type = problems (非选择题 / format: wa): English; no A–E. Same major-question + subpart (and optional nested (i)(ii)) structure; ~10 points per large question unless mirroring a past exam. Use the same [Hard] ~20–25% mix on large questions. Default answers: no: study-guide highlights under each stem block only. answers: yes: append solutions after each question in-file (or user may ask for a separate Solutions note).

• Grading / mistakes — entry-point routing (read first): Before starting any grading, determine which path applies:

After determining path, proceed. Update your Dashboard study-plan section (if maintaining one) when adding new Mistakes files.

错因标签（Error type taxonomy） — shared by MCQ and WA

Every wrong or blank subpart in any Mistakes file must be labelled with exactly one of the following tags in §1 错题速查表 and §2/§2b 解析 header:

§3 定向补强 auto-trigger: only fire for concept, formula, or setup errors (≥2 subparts with the same branchId AND one of these three tags). Do not trigger for calc or misread — those require practice discipline, not concept review.

MCQ mistake log template (…-Problems-Mistakes.md) — required when grading MCQ

Use English ## headings only for the main sections. Four top-level sections (in order):

1. ## 1. Score and summary — 满分、得分、得分率、按题/按小问表、错题速查、🚩收藏速查（表内文字可中英混合）。

   • 错题速查表: wrong/blank subparts — columns: Subpart | 你的作答 | 正确答案 | 错因 tag (one of: concept / formula / setup / calc / misread) | 对应 branchId. The 错因 tag is the agent's best judgment based on the solution; the user may correct it.
   • 🚩 收藏速查表: list every flagged (- [x] 🚩) subpart — columns: Subpart | 是否也错 | 简短备注 (user fills later).

2. ## 2. Wrong / blank — questions, options, and solutions — For each large question that has ≥1 wrong or blank subpart:

   • Paste the full large-question stem (all "You are given" items) so the reader never needs to open the Problems file.
   • Paste each relevant subpart's prompt + all A–E options with legend at the start of §2:
     • Correct: green HTML + ✅ (optional「与你一致」).
     • User wrong: red + ❌.
     • Unanswered: ⚠️; still mark the correct option ✅.
   • Immediately under each subpart block, add #### Solutions — Question n (x) in 中英双语 (bilingual). Language rule for Mistakes §2 / §2b: body text uses Chinese for narrative flow, but every key concept, theorem name, formula label, and technical term must appear in both Chinese and English — e.g. "极大似然估计 (Maximum Likelihood Estimation)", "后验均值 (posterior mean)". English phrasing should be substantive, not just parenthetical glosses — include the English reasoning step or definition when it aids exam recall. All subparts in a displayed question get a solution note (not just the wrong ones): wrong / blank subparts need 【详析】 + 错因 tag; correct subparts get a brief 解题思路 + knowledge recap. 🚩-flagged subparts in this question also get 【🚩见§2b】 cross-reference — their full solution appears in §2b, not here.
   • After all subparts of that question, include a per-question Knowledge highlights blockquote (same format as Problems file but with answers now visible).

   §2 includes correct subparts too: every subpart of a displayed question gets its solution/思路. Subparts that were both correct and un-flagged still appear (brief note is fine).

   §2 density — default vs 精简 (minimal): Default = full stem + all options + all subpart solutions for every large question that has ≥1 wrong/blank subpart. If the user asks for 精简, 轻量, or minimal Mistakes, use compact §2:

   • For large questions that are fully correct and have zero wrong/blank: a one-line summary only (e.g. "Q3: 全对 ✅") — no full restatement.
   • For any large question with ≥1 wrong or blank subpart: keep full stem + all subparts for that question.
   • ## 1, ## 2b, and ## 3 stay full quality.

3. ## 2b. Bookmarked — review notes — Always list every 🚩-flagged subpart here, regardless of whether the parent question appeared in §2 or not. This makes §2b the single authoritative review section for flagged items.

   • For each 🚩-flagged subpart (correct or wrong):
     • Paste the large-question stem ("You are given" block).
     • Paste the subpart prompt + all A–E options (mark correct ✅ + 🚩 badge; add ❌ if also wrong).
     • Write full 解题思路 + Knowledge highlights (same depth as a wrong subpart in §2).
     • Add 错因 tag if it was also wrong.
   • If the user flagged no subparts, this section may say "(No bookmarked subparts this session.)" and be omitted in 精简 mode."

4. ## 3. Key takeaways and further reading — 知识点表（可中英对照列）+ 库内 ASM lesson 链接。Include 🚩-flagged topics in the takeaway table even if they were answered correctly.

   §3 定向补强（auto-trigger）： 触发条件：同一个 branchId（t#-a / t#-b / …）在本套 Mistakes 中出现 ≥ 2 个错因标签为 concept / formula / setup 的错误（calc 和 misread 不计入触发，只统计）。每个达到阈值的 branchId 各写一段 ### 定向补强：【branchId — 主题名】，内容：

   • 错因归类： 列出题号/小问、各自的错因 tag、共性根本原因。
   • 必读梳理： 指向相关附录、Summary、或知识点笔记链接（如 [[T2-1-aggregate-iii-Problems-MCQ]] 附录 A）。
   • 重做清单： 建议立刻重做的题号；做完对照 Key 的 Knowledge highlights 复查。

   若没有任何 branchId 达到阈值，§3 只保留常规知识点表 + 链接（不强制补强段）。若有 calc / misread 错误，在知识点表末尾加一行：「本套有 X 处计算/读题失误，建议下次做题时提高细心度。」（不展开为独立补强段）。

Self-contained principle: The Mistakes file should be readable on its own — the reader should not need to open …-Problems.md to see what a question asked. Every displayed question includes its full original stem (large-question givens + subpart prompts + options).

• Type = mistakes — language: 中英双语 (bilingual). Narrative in Chinese; all key concepts / terms / theorem names in Chinese + English (e.g. "贝叶斯估计 (Bayesian estimation)"). Formulas, definitions, and exam-critical distinctions should carry English phrasing substantial enough for direct exam use — not just a one-word gloss. MCQ 批改遵守本节顺序与「题后紧跟解析」版式。
• After emitting a new Mistakes file with scores: update ASTAM-Practice-Data.json + ASTAM-Practice-Dashboard.md per Practice analytics above.

WA mistake log template (…-Problems-Mistakes.md) — required when grading WA (非选择题)

WA subparts are partial-credit (no single correct letter). Use the same four-section structure as MCQ but adapted for written-answer grading.

Scoring format (§1 scoring table):

Section rules:

1. ## 1. Score and summary — 满分、得分、得分率、上表（按大题汇总）、错题速查（含错因标签）。🚩 收藏速查表 (subpart | 是否也错 | 备注)。
2. ## 2. Wrong / blank — steps, marking, and solutions — For each large question with ≥1 wrong/partial subpart:
   • Paste the full large-question stem (all given items).
   • For each subpart — show in order:
     1. **Marking guide** — list expected steps and their point values (derive from Key if available; otherwise construct from first principles).
     2. **你的作答 vs 标准** — annotate where the approach diverged; assign partial credit per step if applicable.
     3. **错因 tag:** concept / formula / setup / calc / misread.
     4. **【详析】** — full bilingual explanation (same depth as MCQ §2 wrong subparts).
   • After all subparts, include a per-question Knowledge highlights blockquote.
   • Correct subparts get a brief 解题思路 note (no marking guide needed).
3. ## 2b. Bookmarked — review notes — Same rule as MCQ: every 🚩-flagged subpart appears here with full stem + derivation + Knowledge highlights, regardless of correctness.
4. ## 3. Key takeaways and further reading — Same §3 rules as MCQ (知识点表, 定向补强 auto-trigger, same error-tag thresholds).

Key differences from MCQ template: no A–E options; scoring is step-by-step; 错因 tagging is especially important for setup errors (student modeled the wrong quantity).

Knowledge highlights (required for Type = problems)

File division of responsibility:

• This section (SKILL.md): format rules, placement, when to add, language norms.

• reference-knowledge-highlights.md: mandatory block template, T1–T6 topic pitfall bank — read it before generating any T# problems to seed Core / 易错点 bullets.

• reference-astam-generation.md: math alignment examples and extended grading detail.

• Default (answers: no) — MCQ or WA: per large question on …-Problems.md — blockquote Knowledge highlights (study guide — no answers) (MCQ: no correct letters).

• MCQ + Key (user requested …-Problems-Key.md): repeat or expand highlights in the Key after brief rationale.

• WA + answers: yes: after each question’s solution in-file; when grading, highlights appear in Mistakes §2.

After every question (MCQ study guide, WA solution, or Key entry), the blockquote must include:

• First line: > **Knowledge highlights** — for default MCQ stems use > **Knowledge highlights (study guide — no answers)** so spoilers are avoided.
• Core: 1–3 bullets — formulas / steps directly tied to that question
• 易错点 / Common pitfalls: 1–2 bullets — typical exam confusions for that setup

If the user chooses answers: no answers, still add a short Knowledge highlights (study guide) blockquote under the question stem (no full solution). When grading (Type = mistakes), include full highlights in the Mistakes file §2 solutions.

Templates + pitfall bank by T1–T6: reference-knowledge-highlights.md — read it when generating problems so output stays consistent.

Optional: end multi-question files with a small Quick reference card table (see reference file).

Full detail and source list: reference-astam-generation.md.

End-of-task checklist

• Primary T# chosen and reflected in filename (for new files).
• Content grounded in syllabus + reference materials (Past Exams, ASM examples, formula sheets) where relevant.
• Type = problems: each question has Knowledge highlights per reference-knowledge-highlights.md (MCQ default = spoiler-free block on stem file).
• 出题 default: …-Problems.md stems only, answers: no (no solutions / no letters); Key only if requested. Filename uses Problems, not -MCQ.
• Past Exams + ASM examples consulted; syllabus scope for the slice covered; attribution lines present (ASM example ID + Past Exam sitting when adapted); [Hard] on ~1/5–1/4 of large questions (unless user overrides); remaining items Medium+.
• Complexity calibration passed: each large question has scenario unity (cross-subpart dependency); ≥2 non-formulaic subparts; ~1/3 of total subparts are comparison/judgment; draft depth checked against a named Past Exam question.
• New problems or graded Mistakes: ASTAM-Practice-Data.json + ASTAM-Practice-Dashboard.md updated in the same task.
• Type = problems: hidden <!-- astam: t#-x --> (or footer astam-syllabus-map) so Mistakes/ledger can attach branchId quickly.
• Type = problems: every subpart has - [ ] 🚩 on its own line, after subpart question text and before first - [ ] A. option (see canonical placement — no alternatives).
• MCQ numeric subparts: multi-step chains keep ≥3 decimals (or exact storage) for intermediates; final computed first, then options written so one letter is truly correct (not only “closest” due to rounding drift); distractors map to named slips; revise parameters if any tie or ambiguity.
• If a Key exists: spoilers isolated — sibling …-Problems-Key.md, or terminal ## Answer key (spoilers…) block only at file end (no answers interleaved after each question on the practice scroll path).
• Type = mistakes: Mistakes file is self-contained — every displayed question includes full original stem (large-question givens + subpart prompts + options); all subparts get solution notes; 🚩-flagged correct subparts appear in §2 or §2b with full 解题思路.
• User's vault README or Dashboard updated if files changed (if user maintains one).
• Dashboard study-plan section updated with [[wikilinks]] if new/renamed notes should appear in the T1–T6 index or weekly rows (or if the user asked to record progress checkboxes).
• User sees where the file was saved and how it maps to the syllabus.
• MCQ sets: correct A–E labels are spread across subparts (not clustered on one letter); option order re-verified after shuffling. Conceptual subparts: each distractor maps to a named pitfall from the T# pitfall bank.
• Math: stems/solutions/lists/tables use $...$ inline; $$...$$ only as standalone paragraphs; multi-line display uses aligned (or equivalent); no bare \\ stacks without an environment.