無料GitHub で公開

universal-exam-cram-coach-full

帮助学生在临考前进行结构化极速复习：解析课程资料/大纲/重点，按章节生成 wiki 知识库与标准题库，组织针对性刷题与判分，并记录复习进度和错题。当用户即将考试、需要快速复习计划、练习题、错题复盘或考前小抄时使用（关键词：期末/备考/复习/刷题/划重点/错题；exam, cram, study plan, quiz, review）。

不适用于长期学习规划、与考试无关的写作或编程任务。

ZeKaiNie/universal-examprep-skill

インストール方法を見る

含まれるファイル（200）

SKILL.md10.4 KB
.gitattributes1.5 KB
.github/workflows/ci.yml1.9 KB
.gitignore659 B
AGENTS.md38.8 KB
assets/exam-panic.png934.2 KB
benchmark/.gitignore1.2 KB
benchmark/aggregate_matrix.py16.3 KB
benchmark/behavior_smoke/fixtures/mini_course_nostate/references/assets/venn_prompt.svg677 B
benchmark/behavior_smoke/fixtures/mini_course_nostate/references/quiz_bank.json3.2 KB
benchmark/behavior_smoke/fixtures/mini_course_nostate/references/wiki/ch1_stack_queue.md447 B
benchmark/behavior_smoke/fixtures/mini_course_nostate/references/wiki/ch2_trees.md493 B
benchmark/behavior_smoke/fixtures/mini_course_nostate/study_plan.md550 B
benchmark/behavior_smoke/fixtures/mini_course_nostate/study_progress.md429 B
benchmark/behavior_smoke/fixtures/mini_course/references/assets/venn_prompt.svg677 B
benchmark/behavior_smoke/fixtures/mini_course/references/quiz_bank.json4.5 KB
benchmark/behavior_smoke/fixtures/mini_course/references/wiki/ch1_stack_queue.md447 B
benchmark/behavior_smoke/fixtures/mini_course/references/wiki/ch2_trees.md493 B
benchmark/behavior_smoke/fixtures/mini_course/study_plan.md550 B
benchmark/behavior_smoke/fixtures/mini_course/study_progress.md429 B
benchmark/behavior_smoke/fixtures/mini_course/study_state.json390 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_chat_auto_pdf_bad.txt433 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_chat_good.txt462 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_chat_not_persisted_bad.txt186 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_one_shot_good.txt490 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_one_shot_persist_bad.txt605 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_one_shot_probe_error_bad.txt507 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_visual_native_good.txt495 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_visual_preflight_failed_bad.txt487 B
benchmark/behavior_smoke/mock/sample_outputs/artifact_mode_visual_subscription_guess_bad.txt507 B
benchmark/behavior_smoke/mock/sample_outputs/hint_skip_offer.txt349 B
benchmark/behavior_smoke/mock/sample_outputs/language_ask_good.txt461 B
benchmark/behavior_smoke/mock/sample_outputs/language_ask_missing_line.txt270 B
benchmark/behavior_smoke/mock/sample_outputs/language_ask_unglossed.txt211 B
benchmark/behavior_smoke/mock/sample_outputs/language_persist_good.txt216 B
benchmark/behavior_smoke/mock/sample_outputs/language_persist_noncanonical.txt157 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_bilingual_explicit.txt372 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_bilingual_negated.txt306 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_bilingual.txt261 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_question.txt224 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_silent.txt364 B
benchmark/behavior_smoke/mock/sample_outputs/language_urgent_wrong_language.txt262 B
benchmark/behavior_smoke/mock/sample_outputs/notebook_persist_chat_only.txt952 B
benchmark/behavior_smoke/mock/sample_outputs/notebook_persist_good.txt744 B
benchmark/behavior_smoke/mock/sample_outputs/progress_after_confusion.md395 B
benchmark/behavior_smoke/mock/sample_outputs/progress_after_mistake.md403 B
benchmark/behavior_smoke/mock/sample_outputs/provenance_answer.txt492 B
benchmark/behavior_smoke/mock/sample_outputs/quiz_output_good.txt351 B
benchmark/behavior_smoke/mock/sample_outputs/quiz_output_invented.txt317 B
benchmark/behavior_smoke/mock/sample_outputs/resume_message.txt224 B
benchmark/behavior_smoke/mock/sample_outputs/scope_override_good.txt258 B
benchmark/behavior_smoke/mock/sample_outputs/scope_override_missing.txt127 B
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_ai_answer_good.txt1.2 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_formula_first.txt1.1 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_good.txt1.2 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_liberal_good.txt1.4 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_missing_warn.txt608 B
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_no_source.txt1.1 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_optin_closers.txt1.5 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_skip_ask.txt948 B
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_unlabeled_source.txt517 B
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_unsolicited_closers.txt1.5 KB
benchmark/behavior_smoke/mock/sample_outputs/teaching_template_warn_title.txt581 B
benchmark/behavior_smoke/mock/sample_outputs/time_budget_1day_bad.txt340 B
benchmark/behavior_smoke/mock/sample_outputs/time_budget_1day_bank_checkpoint.txt341 B
benchmark/behavior_smoke/mock/sample_outputs/time_budget_1day_good.txt695 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_answer_after_prompt.txt264 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_answer_before_prompt.txt264 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_answer_side_first.txt261 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_answer_text_before_prompt.txt243 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_good.txt232 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_missing_asset.txt182 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_path_only.txt163 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_prose_before_image.txt234 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_question_label_late_asset.txt250 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_unlabeled_solution_before_prompt.txt233 B
benchmark/behavior_smoke/mock/sample_outputs/visual_first_unsafe_url.txt178 B
benchmark/behavior_smoke/mock/sample_outputs/window_recheck_bad.txt346 B
benchmark/behavior_smoke/mock/sample_outputs/window_recheck_good.txt665 B
benchmark/behavior_smoke/mock/sample_outputs/window_recheck_recall_only_bad.txt333 B
benchmark/behavior_smoke/mock/sample_outputs/window_recheck_test_good.txt580 B
benchmark/behavior_smoke/mock/sample_outputs/workspace_confirm_good.txt797 B
benchmark/behavior_smoke/mock/sample_outputs/workspace_confirm_silent.txt453 B
benchmark/behavior_smoke/mock/sample_outputs/zero_basic_explain.txt1.2 KB
benchmark/behavior_smoke/mock/sample_outputs/zero_basic_legacy_only.txt442 B
benchmark/behavior_smoke/mock/sample_transcript.jsonl517 B
benchmark/behavior_smoke/mock/state_after_confusion.json203 B
benchmark/behavior_smoke/mock/state_after_mistake.json215 B
benchmark/behavior_smoke/README.en.md9.9 KB
benchmark/behavior_smoke/README.md7.4 KB
benchmark/behavior_smoke/run_behavior_smoke.py97.2 KB
benchmark/behavior_smoke/scenarios.json12.7 KB
benchmark/calibrate_matrix.py20.0 KB
benchmark/calibrate.py7.3 KB
benchmark/calibration/kappa_n24_summary.json1.5 KB
benchmark/calibration/README.en.md3.2 KB
benchmark/calibration/README.md2.6 KB
benchmark/config.example.json431 B
benchmark/config.matrix.example.json1.1 KB
benchmark/docs/coverage-matrix.en.md17.6 KB
benchmark/docs/coverage-matrix.md15.2 KB
benchmark/docs/img/hard_algo_correct_en.svg2.7 KB
benchmark/docs/img/hard_algo_correct_zh.svg2.7 KB
benchmark/docs/img/hard_algo_hallu_en.svg2.7 KB
benchmark/docs/img/hard_algo_hallu_zh.svg2.7 KB
benchmark/docs/img/hard_psyc_correct_en.svg2.7 KB
benchmark/docs/img/hard_psyc_correct_zh.svg2.7 KB
benchmark/docs/img/hard_psyc_hallu_en.svg2.7 KB
benchmark/docs/img/hard_psyc_hallu_zh.svg2.7 KB
benchmark/docs/img/oos_algo_abst_en.svg2.7 KB
benchmark/docs/img/oos_algo_abst_zh.svg2.7 KB
benchmark/docs/img/oos_psyc_abst_en.svg2.7 KB
benchmark/docs/img/oos_psyc_abst_zh.svg2.7 KB
benchmark/docs/img/realexam_algo_correct_en.svg2.7 KB
benchmark/docs/img/realexam_algo_correct_zh.svg2.7 KB
benchmark/docs/judge-calibration.en.md13.1 KB
benchmark/docs/judge-calibration.md9.9 KB
benchmark/docs/materials_sources.en.md3.2 KB
benchmark/docs/materials_sources.md2.7 KB
benchmark/docs/matrix_pipeline.en.md9.5 KB
benchmark/docs/matrix_pipeline.md8.4 KB
benchmark/docs/related_benchmarks.en.md5.5 KB
benchmark/docs/related_benchmarks.md4.6 KB
benchmark/docs/running-real-runs.en.md10.0 KB
benchmark/docs/running-real-runs.md8.0 KB
benchmark/docs/test_tiers.en.md7.7 KB
benchmark/docs/test_tiers.md6.8 KB
benchmark/docs/testing-audit.en.md17.2 KB
benchmark/docs/testing-audit.md14.6 KB
benchmark/docs/testing-walkthrough.en.md23.5 KB
benchmark/docs/测试流程详解.md19.0 KB
benchmark/drift/convert_session_log.py12.9 KB
benchmark/drift/docs/live_agent_pilot.md6.6 KB
benchmark/drift/fixtures/live_logs/good_session.expected.jsonl1.5 KB
benchmark/drift/fixtures/live_logs/good_session.md1.6 KB
benchmark/drift/fixtures/live_logs/live_smoke_golden.jsonl2.2 KB
benchmark/drift/fixtures/live_logs/live_smoke_golden.md2.3 KB
benchmark/drift/fixtures/mini_course_long_state/references/quiz_bank.json1.3 KB
benchmark/drift/fixtures/mini_course_long_state/references/wiki/ch1_stack_queue.md541 B
benchmark/drift/fixtures/mini_course_long_state/references/wiki/ch2_trees.md413 B
benchmark/drift/fixtures/mini_course_long_state/study_plan.md426 B
benchmark/drift/fixtures/mini_course_long_state/study_progress.final.bad.md151 B
benchmark/drift/fixtures/mini_course_long_state/study_progress.final.good.md272 B
benchmark/drift/fixtures/mini_course_long_state/study_progress.initial.md151 B
benchmark/drift/fixtures/mini_course_long_state/study_state.json145 B
benchmark/drift/fixtures/mini_course_long/references/quiz_bank.json1.3 KB
benchmark/drift/fixtures/mini_course_long/references/wiki/ch1_stack_queue.md541 B
benchmark/drift/fixtures/mini_course_long/references/wiki/ch2_trees.md413 B
benchmark/drift/fixtures/mini_course_long/study_plan.md426 B
benchmark/drift/fixtures/mini_course_long/study_progress.final.bad.md151 B
benchmark/drift/fixtures/mini_course_long/study_progress.final.good.md272 B
benchmark/drift/fixtures/mini_course_long/study_progress.initial.md151 B
benchmark/drift/README.en.md14.6 KB
benchmark/drift/README.md12.3 KB
benchmark/drift/run_drift.py79.7 KB
benchmark/drift/run_live_smoke.py26.4 KB
benchmark/drift/scenarios/live_smoke_basic.json681 B
benchmark/drift/scenarios/long_session_basic.json793 B
benchmark/drift/scenarios/long_session_state.json994 B
benchmark/drift/scenarios/mode_urgent_no_questions.json1.1 KB
benchmark/drift/scenarios/window_persist.json1.0 KB
benchmark/drift/templates/live_session_template.md1.3 KB
benchmark/drift/templates/live_smoke_turns.json1.6 KB
benchmark/drift/transcripts/bad_checkpoint_reset.jsonl877 B
benchmark/drift/transcripts/bad_goal_drift.jsonl226 B
benchmark/drift/transcripts/bad_plan_drift.jsonl729 B
benchmark/drift/transcripts/bad_progress_loss.jsonl1013 B
benchmark/drift/transcripts/bad_provenance_drift.jsonl583 B
benchmark/drift/transcripts/bad_quiz_invention.jsonl590 B
benchmark/drift/transcripts/bad_urgent_1day_questions.jsonl840 B
benchmark/drift/transcripts/bad_wiki_overread.jsonl430 B
benchmark/drift/transcripts/good_session_state.jsonl7.0 KB
benchmark/drift/transcripts/good_session_urgent_1day.jsonl2.8 KB
benchmark/drift/transcripts/good_session.jsonl5.5 KB
benchmark/drift/transcripts/window_persist_session.jsonl7.3 KB
benchmark/fixtures/mini_course_matrix/.gitignore250 B
benchmark/fixtures/mini_course_matrix/config.json394 B
benchmark/fixtures/mini_course_matrix/items.jsonl1.6 KB
benchmark/fixtures/mini_course_matrix/materials/_combined.txt502 B
benchmark/fixtures/mini_course_matrix/raw_ws/lecture01.md379 B
benchmark/fixtures/mini_course_matrix/skill_ws/references/wiki/ch01_minios.md485 B
benchmark/gen.py13.0 KB
benchmark/items/items_algo_full_q.jsonl13.5 KB
benchmark/items/items_pilot_6006.jsonl4.5 KB
benchmark/items/items_pilot_psyc110.jsonl4.5 KB
benchmark/items/items_psyc_full_q.jsonl7.5 KB
benchmark/items/items.example.jsonl2.0 KB
benchmark/items/README.en.md3.3 KB
benchmark/items/README.md2.6 KB
benchmark/judge.py23.8 KB
benchmark/materials/README.en.md2.7 KB
benchmark/materials/README.md2.4 KB
benchmark/README.en.md12.7 KB
benchmark/README.md10.8 KB
benchmark/rejudge.py18.2 KB
benchmark/report_matrix.py29.1 KB
benchmark/REPORT.en.md7.1 KB
benchmark/REPORT.md6.4 KB
benchmark/report.py22.9 KB
benchmark/results/chart_delta_en.svg1.5 KB

SKILL.md（原文）

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Universal Exam Cram Coach — Root Router

This language-neutral router dispatches last-minute exam prep to the chapter-wiki, bank-only, persistent control layer and its wording packs; it is not a duplicate manual.

Language dispatch

Read the canonical, language-neutral study_state.json.language code and load the matching compatibility entry plus its per-skill wording pack BEFORE emitting any student-visible output:

zh (display choice 中文) → locales/zh/SKILL.md plus the selected sub-skill's zh wording pack under locales/zh/skills/
en (display choice English) → locales/en/SKILL.md plus the selected sub-skill's en wording pack under locales/en/skills/
bilingual (display choice 双语) → compose the zh and en wording block by block, with zh first and a > EN: mirror for each block (composition rules in docs/language-policy.md)

中文, English, and 双语 remain accepted user-facing input aliases. On first contact one combined ask sets mode, budget, and language, then show the independent material-processing choice 轻量按需（推荐） / 完整建库; exam_start.py confirm persists them with the exact workspace/materials receipt. Missing, urgent, accepted-default, and legacy processing choices mean lightweight; only explicit full opens complete ingestion. A later reconfirm with no processing flag preserves an existing canonical choice. Later update_progress.py set --language applies next turn. Default English unless the student opened in Chinese; bilingual is explicit-only.

Control layer (behavior)

Behavior lives in skills/exam-cram/SKILL.md and these subskills:

Sub-skill	Role
`exam-ingest`	Build/validate workspace
`exam-tutor`	Lazy chapter teaching
`exam-study-guide`	Typed guide and visual artifact gate
`exam-quiz`	Bank-only selection/grading
`exam-review`	Replay mistakes/confusions
`exam-cheatsheet`	Final handout
`exam-audit`	Read-only workspace health check
`exam-help`	Quick reference
`confusion-tracker`	Concept-confusion tracking

Generic-agent fallback: AGENTS.md.

Install & run essentials

Under scripts/, use exam_start.py status, then exam_start.py confirm --course <name> --materials <dir> --workspace <ws> --mode <mode> --time-budget <tier> --language <lang> --processing-mode <lightweight|full>. It writes the confirmation/state/runtime receipt. Default lightweight_session.py inventories names and processes only current-phase PDF pages or definitely single-frame PNG/JPEG/BMP sources through host-native vision: at most eight primary pages and one active batch. A single page uses no contact sheet; multi-page overview sheets partition primary pages in groups of at most four at roughly 768 px per tile. New schema-3 visual receipts require the generic component token strategy and enumerate stable teaching-item IDs plus generic text|figure|mixed prompt/answer components. A cross-page item repeats on each page that supplies one of its prompt components, with exact page↔component coverage. Detail calls may combine only same-target prompt components, solution calls only same-target answer components, and every component crop receives a separate semantic review that detects exactly its declared target/context IDs with no unrelated content or student attempt. Only prompt components may be context-only; every answer component contains its target. Page answer provenance prevents student attempts or unknown pages from masquerading as official solutions, and every registered official-solution page must contribute an answer component. Additive register-answer-dependency binds exact answer-locator pages; planned batches may auditably replace/narrow or remove a binding with set-answer-dependency / remove-answer-dependency. All canonical visible evidence is PNG under .lightweight/assets/, with exact model-input receipts and hash/magic/dimension checks. Schema-2 visual receipts and the legacy figure-only token strategy remain read-only history; any legacy-strategy active attempt is restricted to status or auditable abandon and cannot silently become schema 3. An unfinished planned/visual-ready batch may close only with receipt-backed abandon --reason; replace-taught --reason preserves a taught predecessor/event as superseded history, revalidates its dependency revisions, and plans an exact-slice successor with the same dependency pages. After an unabridged walkthrough, mark-taught --taught-item-ids <exact IDs> binds notebook/chNN.md#anchor, distinguishes inspected pages from taught items, and recoverably publishes phase_evidence.lightweight_batches; only current unsuperseded attempts enter the completion denominator. Routine status is generation-stable and read-only; validation checks metadata plus physical identity only. Exact hashes are reserved for state transitions, completion, or explicit status --verify-live. Lightweight verified additionally requires two revision-bound checkpoints, including one pass, from the immutable stat-only baseline of a quiz bank that pre-existed initialization. It runs no full ingestion, Study Guide, or PDF. Explicit full opens ingest_course.py; the orchestrator and lower-level workspace builder/compiler all enforce the same exact-pair/runtime/choices/full gate. Exit 10 routes to typed ingest_review.py. update_progress.py owns study_state.json; study_progress.md is generated. Official selectors are select_questions.py / select_hard_questions.py.
Workspace file contract (wiki / quiz bank / state / asset metadata): docs/file-format.md.
Language policy (single-language purity, EN canonical vocabulary, persisted canonical values): docs/language-policy.md.
Host loading: docs/agent-portability.md.
PDF capabilities differ by host; use the audited, no-silent-download routing table in docs/pdf-capability-adapters.md.
Missing/legacy artifact_mode is chat; explicit standing visual or a one-shot chapter artifact invokes exam-study-guide, while cheat-sheet PDF uses exam-cheatsheet. An ambiguous PDF request asks which once. Never infer subscription. Persist with update_progress.py set --artifact-mode chat|visual.
processing_mode and artifact_mode are independent. Lightweight never generates a Study Guide; a saved visual preference remains dormant and effective output stays chat until explicit full. Full does not imply a PDF. MinerU, Docling, and LangGraph are explicit-named-request, remote/cloud-host-only capabilities and are never probed, downloaded, installed, imported, executed, or accepted as callable local runners.
preferences.interaction_style stores only batch|step_by_step. A stored step-by-step choice is effective only in full with no_questions=false; otherwise it is retained but dormant and effective cadence is batch. In effective step mode, select the first pending teaching_examples.json item from one locked snapshot and persist it through the marker-bound record-taught-example path. Existing unbound teaching IDs are valid batch history; a bound ID carries exact notebook-block and manifest-item hashes that remain live-validated after cadence changes. Guide publication preserves valid bound blocks and rejects stale or unbound markers. Every teaching-baseline ID must still have a current teaching-manifest snapshot; a quiz-only copy is insufficient.
In an ingestion-v2 structured workspace, answer_explanation_mode is independent from processing/artifact mode. Its stored-schema fallback is ordinary, but full-v2 Guide entry must first perform a native-child capability handshake. When the host can prove a fresh independent child context per item and can restrict its input and tools to that exact item, default to isolated unless the user opted out; persist the mode, notify once about extra host quota/time, and require no second API key or external-upload consent. Otherwise stay ordinary and explain the limitation. Both routes run study_guide_author.py prepare, fill fixed annotations, require one detailed beginner-first explanation per item, persist notebooks, compile, create/attach/verify claims, and import the canonical full Guide. In ordinary, the annotation contains the explanation with ai_supplement provenance and claims no isolation. In isolated, each fresh/stateless tool-disabled invocation sees only the fixed question, official answer when present, target language, and target-scoped assets; it returns answer_explanation plus non-rendered coverage and is imported with a separate host-owned receipt. A separately billed external Provider is an explicit-user-request fallback only and retains no-upload planning plus exact-plan pricing/privacy/upload consent. A model family, subscription, API key, full, or visual alone never proves native isolation. Target-scoped means target_item_only, or prompt-only target_with_required_context with exact sorted required_context_ids; answer assets remain target-only. Packet, annotations, notebook bindings, manifest, rendering and QA all bind the chosen mode. A language/mode/fact/asset change makes the chain stale; only isolated reruns the per-item receipt chain. New v2 Guides omit generic self-check panels. A hand-written complete v2 Guide draft is a no-Python-only, unverified fallback. Ingestion-v1 remains read-only and cannot claim current v2 gates.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

ログインしてレビューを書く

同じリポジトリのスキル

概要と使いどころ