Teach exactly one current wiki chapter, using metaphors and formula dissection. In zero-basic mode, explain every linked key question with the fixed seven-step walkthrough. Run algorithms before rendering diagrams. This skill teaches; exam-quiz alone quizzes and scores.
-
Load one slice. Read study_state.json.processing_mode first. In
lightweight, call lightweight_session.py status, plan only the current
source/page range if it is not already planned, visually inspect those pages,
and import the generic item/component manifest with record-visual; teach only a
schema-3 visual_ready batch. A schema-2 visual_ready receipt is quarantined
read-only: auditably abandon it and plan a new attempt, never teach from or
silently upgrade it. While still planned, keep register-answer-dependency
additive; use set-answer-dependency --reason to replace/narrow exact answer
pages and remove-answer-dependency --reason to remove them. Do not call
ingestion/OCR, preload later pages, or require a
wiki. In full, read exactly one current references/wiki/chN_*.md. A missing
full-mode file means abstain, name it, and never improvise. If full-mode teaching
examples exist, run python "${CLAUDE_SKILL_DIR}/scripts/list_teaching_examples.py" --workspace <ws> --chapter <N> --json and use only its returned slice. When the
full-mode effective cadence below is step_by_step, use --next-pending instead
of loading the whole chapter example slice. A nonzero
exit is an invalid/unreadable inventory, not “no examples”; report it.
-
Teach reproducibly. Give each concept one concrete metaphor. For STEM, state every formula symbol and unit, then one small hand-computable example. Persist math as $...$ or $$...$$; never leave raw \frac, \sum, or other TeX as the final reading view.
-
Use every walkthrough block in order for every stored/teacher-flagged question and every linked question in zero-basic mode.
Full-mode pacing: read the stored preference plus its reported effective and
dormant state. study_state.json.preferences.interaction_style stores only
batch|step_by_step; missing legacy state means batch. This optional preference
is independent from processing_mode, artifact_mode, and
answer_explanation_mode, and is not a fourth required startup choice. Persist an
explicit change only with update_progress.py --workspace <ws> set --interaction-style <batch|step_by_step> (or the strictly validated canonical
--pref interaction_style=...). It never changes the lightweight page-batch route.
This option applies only to full-mode teaching_examples.json items. It does not
claim coverage of the chapter bank, typed question units, or the lightweight
page-batch route.
- Effective
batch: use the normal full-mode flow. A true
preferences.no_questions=true or any non-full processing mode makes a stored
step_by_step choice dormant without overwriting it. A stored batch choice
remains ordinary batch cadence.
- Effective
step_by_step: call list_teaching_examples.py --workspace <ws> --chapter <N> --next-pending --json. It requires processing_mode=full,
no_questions=false, exact current_phase, and valid scoped manifest/state data.
It reads the manifest, state, notebook bindings, and baseline within one
consistent workspace lock, then returns the first manifest-ordered pending item.
A missing manifest, malformed state, or nonzero selector exit blocks the pacing
decision; report it and do not guess another item. Two bindings may not share one
notebook_ref. Only a missing notebook entry or anchor/marker/hash/revision drift
may return to pending with bounded stable diagnostics. Link/reparse topology,
non-directory/non-regular targets, path escape, invalid UTF-8, an unterminated
fence, parse/block corruption, schema/scope/baseline damage, duplicate evidence,
and unexpected_evidence are fatal.
Unbound IDs already present in phase_evidence[N].teaching_examples are legal
batch/legacy history rather than corrupt step evidence; any ID with a
teaching_example_bindings record must pass its live notebook-block and
manifest-item hash checks regardless of the currently selected cadence. Teach
exactly that one item this turn, but complete all seven blocks below; never split
one walkthrough across turns. Do not infer progress from notebook presence,
language-specific prose, or “I understand” / . If ,
means only that this full teaching roster
has no pending item, including an empty roster; it never completes the chapter or
bypasses Guide, bank, typed-unit, asset, checkpoint, or phase gates.
A structurally sound current roster with either a stale manifest/notebook binding
or an append-only newly added item is a named mount warning so
manifest-order re-teaching remains legal. Structural/scope/baseline corruption
stays ; the old Guide/completion receipt remains ineligible. Teaching IDs
use the shared 1–200-character Guide-safe Unicode contract; keep an incompatible
source-facing label in source/title metadata instead of changing a stable ID. If
the ID alone produces an empty Markdown slug, the notebook entry needs a
descriptive title. Every retained baseline ID must have a current teaching
snapshot in the same canonical chapter under exact ; a
quiz-only copy cannot substitute.
-
Show question assets first. Before explaining, hinting, or solving any stored question with requires_assets=true or maybe_requires_assets=true, render every question-side question_context / figure / diagram / table asset, labelled 题面图 or Question-side asset. Only afterward may solution/review show official answer_context / worked_solution, labelled 答案图 or Answer-side asset. Preserve but do not display or teach from student_attempt; it is neither prompt nor official/material answer evidence. Treat its physical path as globally tainted across quiz, teaching, and all content units, folding safe slash/backslash aliases and Windows case aliases; never display an official declaration of that path. Reject same-item prompt/answer reuse. Cross-item official prompt/answer reuse without an attempt is legal, and distinct official plus attempt paths remain usable. Missing/unreadable files block a structured workspace and return to validation/exam-ingest; a UI that cannot render the existing image must skip the item. A path is not an image. Prefer python <package-root>/scripts/show_question_assets.py --workspace <ws> --id <qid> --lang <zh|en>; exit 1 means skip. Apply the same gate to stub / page_reference prompts.
In lightweight schema 3, apply this rule to generic components rather than only figure questions. Use the item's text|figure|mixed kind honestly; show every prompt component required to understand the target before teaching, including declared shared context, and never display an answer component until solution/review. A detail call may combine prompt components only for the same target. Trust a component only after its separate crop review detects exactly allowed_detected_item_ids (target plus all declared contexts, or a declared non-empty context-only crop) with no unrelated content or student attempt. A text-only prompt may use a cross-file official answer without being relabelled as a figure item; only official_solution parent pages may provide answer components, and every registered official page must be covered.
-
Run diagram algorithms first. For trees, traversals, graphs, and state machines, actually run the standard Python algorithm before rendering. State that textbook conventions apply and teacher-specific rules prevail. Without Python, show the textual/ASCII/Mermaid derivation and label it 「未经程序验证」.
-
Track state and provenance. Mark material, AI supplement, and AI-generated answers with the canonical labels above. Why/what/how-derived follow-ups invoke confusion-tracker and python "${CLAUDE_SKILL_DIR}/scripts/update_progress.py" --workspace <ws> add-confusion; initialize missing state when Python works.
-
Record evidence; complete only through the gate. Use record-phase-evidence for wiki, visual, notebook, and bank checkpoint evidence (--kind checkpoint --ref <qid> --outcome passed|wrong|skipped). Batch-mode full teaching examples may use its ordinary teaching-example kind, producing legitimate unbound history; effective full step_by_step must instead use the marker-bound record-taught-example path above. Bound history remains live-validated after switching back to batch. Every ID retained by teaching_baseline.json must have a current teaching_examples.json snapshot; a matching quiz item alone cannot satisfy or exhaust the teaching roster. verified requires at least two handled bank items and one pass. set --phase <N> is only explicit navigation/repair, never completion.
In lightweight, never invoke exam-study-guide; after persisting the full
walkthrough and updating progress, bind the batch with
lightweight_session.py mark-taught --batch-id <id> --notebook-entry <path> --taught-item-ids <exact-comma-separated-IDs-from-the-visual-receipt>. The
inspected page list is context, not proof that every item on those pages was
taught; close only the exact item IDs enumerated during visual review.
Plan the next pages only when the learner reaches them. Without a pre-existing
standard bank, no verified checkpoint exists and completion is capped at
covered_unverified. In full, after all current-chapter material has persisted
walkthroughs, invoke exam-study-guide to build, validate, and import the
profile=full notebook/chNN.guide.json. Its de-duplicated teaching-example +
all-bank + typed-question denominator is a coverage gate, with gradable=false
bank records retained as teaching-only Guide content; it is not proof of semantic
recall. Effective missing/unknown artifact_mode is chat: typed import is
enough before complete-phase --status covered_unverified|verified, with no
HTML/PDF. Standing visual must also select the PDF route, render, bind receipts,
accept every page, and reach artifact_ready=ready. A one-shot artifact request
temporarily overrides chat without changing the standing value. Never infer a
subscription or install dependencies silently. Language changes stale the
manifest/artifact: route to exam-study-guide for relocalization, refreshed
claims/receipt, re-import, rerender, and repeat QA. A request for “all examples”
remains under ; time pressure may shorten prose, not omit
required items or language blocks.
-
Apply the time tier. Read mode and budget from state:
≤1天: no opening preference or reflective follow-up; teach now. This does not ban bank-backed drills or checkpoints. Explicit 「不要出题 / 不要问我」 persists no_questions=true, emits no interactive question, and caps completion at covered_unverified.1-3天: occasionally recheck earlier difficult/confused points and reteach forgotten ones.3-7天: add taught points to the knowledge window; ask whether an out-of-window point is remembered before restoring it.>7天: test an out-of-window point with its linked hard bank item; pass → window-set-status ... --status 已实测, fail → reteach. A point/index locator is required; add chapter for ambiguous names.
