Assistant-Authored Study Drills

Use this skill when the user wants practice questions you write rather than official program items. It owns the drill orchestration — not the rules that other skills already own.

This skill owns orchestration; point elsewhere for the rest

• Item quality (strong distractors, key-position spread, cold-read resistance, stem–option shape) → the agonara-student "MCQ item quality" bar. Hold every item you author to it; do not restate it here.
• Button delivery (clarify continuity, button-label-is-option-text, immediate user_response, timeout recovery) → button-quiz-loop. Run every button drill under it.
• Oral / free-response / case drills → oral-exam-drills. Use it when the drill is not single-select.
• Per-item assess-then-advance loop (record the key + each distractor's mistake privately, gate each item, give the required verdict before advancing) → the agonara-student "Running an authored MCQ practice quiz" loop. Run every authored item under it; don't restate it here.

This skill is what connects them into a coherent authored-MCQ session.

Trigger

• The user asks to be "drilled" on a topic.
• The user wants non-official practice before or instead of official checks.
• The user asks for a block of ordinary questions followed by a harder/craftier block.

Core rules

1. Provenance is explicit. Say plainly that these are your practice questions, not official Agonara items — never let it stay ambiguous.
2. One item at a time unless the user explicitly asks for batch delivery.
3. Acknowledge the commentary, not just the chosen option. If they answer with a choice plus reasoning, respond to the reasoning briefly.
4. Respect sequencing literally. If the user says "1/2/3 in order, not in one message" or "wait for «дальше»", stop after the requested slice and wait.
5. Assess before advancing. The verdict's content is the agonara-student assess-then-advance rule (required, not optional) — don't shrink it to a bare "right/wrong," and don't expand it into a mini-lecture.

Established cadence beats generic defaults

If a drill has already settled into a working rhythm — answer → verdict → next question — preserve it unless the user asks to change it.

• Do not drop the verdict just because you switched to button mode. To keep the button chain unbroken, place it at the start of the next clarify prompt (see button-quiz-loop) — the verdict is still the full agonara-student one, just bundled into the next card.
• If you changed the scheme mid-run by accident, say so, restore the prior format, and restart the affected item.

Difficulty shaping

A good sequence usually has two phases:

1. Standard understanding — establish the main distinctions and vocabulary.
2. Cunning / tricky — force fine-grained distinctions between near-neighbours: all options defensible at first glance, the key the most precise rather than the most slogan-like, targeting boundary confusions and overgeneralizations.

For a "really hard" follow-up block, shift away from term-definition recognition toward near-neighbour case comparisons where 2–3 options are defensible and differ only in the governing principle.

Coverage when the request is "the whole corpus"

If the user asks to cover all topics / the whole syllabus:

1. build a topic map first (visible, or at least internal);
2. distribute items across the syllabus before deepening one fertile subtopic;
3. do not let a quality fix silently collapse coverage into the easiest-to-write niche;
4. if the user says coverage narrowed, acknowledge it and rebalance the remaining set.

Two-contour balance: meaning and course nomenclature

Coverage above spreads items across the syllabus by topic. This is the second
coverage axis — across the cognitive layer — and it is just as load-bearing.
Understanding alone is not test-resistance: a learner strong on principles, cases,
and anti-stretch reasoning can still lose real test points on the nomenclature
layer (who distinguished what; how an author's classification is named and built;
named categories, types, pairs, lists). A run that trains only the comfortable
analytical contour and skips the factual one leaves that gap intact — the
documented failure mode, not a hypothetical.

• Train both contours, deliberately. A substantial run must exercise meaning
  and exact course fact, mirroring how the course actually tests — not whichever
  is more elegant to write. If the course checks fixtures (names, typologies,
  categories, author schemes, precise content details), the run checks them too.
• Required item-type mix. Across a run, include: principles/distinctions;
  cases/application; authors/typologies/categories/named schemes; and exact details
  from the materials. Don't let the set collapse into the analytical bucket.
• No drift to the comfortable side. Catch yourself defaulting to "meaning"
  items because they read better. If you notice the mix has skewed, say so and
  rebalance — same discipline as the topic-coverage rebalance above.
• Prioritise known-weak fixture zones. Weight the factual contour toward the
  zones the learner has actually slipped on — named classifications, typologies,
  categories, author breakdowns, exact course content — over zones already solid.
  Use learner-state weak areas and prior-session history as the priority signal.
• Fixture distractors compete on the fixture. For nomenclature items,
  distractors are adjacent author schemes, neighbouring categories, and near-miss
  formulations — the same confusions the real test exploits. This is the
  agonara-student item-quality bar applied to the factual layer (a real nameable
  mistake per distractor); don't restate the bar, instantiate it.
• Verdict names the error type. Fold into the required agonara-student
  assess-then-advance verdict which contour the miss sits in: a logic/reasoning
  error, a plain fact error, a confusion of adjacent categories, or not knowing a
  specific author/scheme. This is a refinement of the verdict's "misconception each
  wrong option encodes," not a replacement for it.

One line: prepare the learner not just to understand the course but to survive
how the course is actually tested.

Source-grounded drills

When the user asks for questions "based on the primary source," treat it as a grounding requirement:

• read the original lecture/material/document first;
• build each hinge from an exact distinction, caveat, hierarchy, exception, or formulation in that source;
• make distractors compete by staying close to the source (nearby formulations, adjacent categories, tempting misreadings of the same paragraph);
• when correcting, name the source-level hinge briefly;
• do not imply documentary grounding unless you actually inspected the source this session.

Teach-then-drill contract

When the user asks you to teach from the source first, then quiz, the teaching recap is the contract for the first wave:

• do not test a subtopic you did not introduce in that recap; either add it to the recap first or say explicitly you are expanding beyond it;
• if a section is too large to cover fully, mark what was covered and what was deferred;
• if the user says "that wasn't in your summary," treat it as a valid sequencing complaint, not nitpicking — "it exists in the source" is not a defense.

Live quality repair

When the user flags an item as weak, leaky, or mismatched — distractors too easy, the key too visible, options that don't answer the stem, or correct answers clumping in one slot:

1. agree plainly if the criticism is right; do not defend the item or count it as a clean solve;
2. name what leaked (e.g. one option markedly more precise, distractors too distant, position patterning);
3. void the item and reissue a stronger version on the spot;
4. tighten the remaining set so the same flaw does not recur.

This is the repair workflow; for what makes an item strong in the first place, hold to the agonara-student item-quality bar.

Underdetermined-case handling

If the user points out a case is under-specified or supports no single determinate answer:

• treat it as a potentially strong analytical observation, not resistance;
• separate what follows from the stated facts from what needs extra assumptions;
• do not "repair" by re-asking about the same factual setup unless they want that;
• reward correct identification of the ambiguity when they still explain the conditional branches.

Capability questions, answered first

If the user asks whether you can use native Telegram polls or drive an external bot for the quiz, answer the capability directly and first (native polls do not work in private 1-on-1 chats; you do not operate other bots). Offer a short alternative only after the boundary answer — do not turn it into a long workaround detour.

Ending the set

A drill ends; it doesn't trail off.

• Planned set complete: close with a short summary — the user's strongest distinctions and remaining weak spots — not another question.
• Early exit: if the user stops mid-set, give that same summary for what was covered rather than dropping the thread.
• In button delivery, this summary is the terminal clarify-exit (the final wrap-up in button-quiz-loop), not a mid-chain message.

Pitfalls

• Letting an authored drill blur into official assessment handling.
• Restating the item-quality or delivery rules here instead of pointing to their owners (item quality lives in agonara-student).
• Ignoring the user's meta-feedback that an item is weak or leaky — repair it (above) instead of defending it.
• Advancing past a "wait for «дальше»"-style cue.

References

• references/user-calibration-notes.md — calibration cues from prior sessions.
• references/user-specific-preferences.md — what this user rejects in practice-question design.
• references/button-flow-failures.md — concrete button-quiz failure modes and recovery for this user (delivery detail; the canonical rules live in button-quiz-loop).