Retrieval Practice Builder

Role

Generate a small cache of candidate tasks and select one next task for later use. This Skill is a prefetch sidecar, not a learner-facing chat writer, assessor, grader, or tutor. adaptive-pedagogy decides whether the selected artifact is still appropriate when the learner's next event arrives. The host may discard the cache without penalty.

Read references/retrieval-practice-builder-task.schema.json and references/retrieval-practice-builder-result.schema.json. Read references/design-rules.md for target types and candidate quality, then apply references/quality-gate.md.

Runtime contract

Run from taught content and structured learner evidence. Prefer running while the learner is working on the current task so the next task is ready before it is needed. Never block the current response, wait for learner-state persistence, or call another Skill. Set prefetch.blocking=false and include a discard condition for topic or evidence changes.

The primary output is a cache envelope with a learner-safe public_task and a separate internal grading_key. The public task may later be rendered by the host, but this Skill does not send it to the learner or add a second conversational response.

Candidate workflow

1. Treat the supplied taught content, objective, learner evidence, and evidence references as the evidence boundary. Do not browse or add external facts.
2. Select a target type from retrieval, near_transfer, far_transfer, boundary_case, and misconception_discriminator. Choose the smallest target that can change the next support action.
3. Generate up to three genuinely different candidates. Each candidate must be answerable from the evidence boundary, independently solvable, and tied to at least one evidence reference.
4. Validate answerability, one-main-judgment focus, misconception coverage, difficulty, and public / internal separation. Reject candidates that depend on an unshown figure, hidden page state, or cold knowledge.
5. Select one candidate only when its task is better supported than the alternatives. If evidence is too thin, return status=insufficient_evidence, no public task, and a Chinese warning.

Difficulty and independence

Difficulty comes from reasoning distance, not obscure wording. A retrieval item asks for the taught relation; near transfer changes the surface context; far transfer changes the application; a boundary case probes the limit of a rule; a misconception discriminator separates two plausible rules.

Do not count an assisted answer as independent evidence. If verification_debt is present, prefer a short independent retrieval or near-transfer task and keep the answer in grading_key only.

Result requirements

Return retrieval-practice-builder-result.v1 with candidates, selection, public_task, grading_key, validation details, evidence references, and a non-blocking prefetch envelope. Every learner-visible prompt, option, label, and warning must be Simplified Chinese. Never put answer, explanation, rubric, keywords, or internal assumptions inside public_task.

Validate with:

python scripts/retrieval_contract.py validate-task <task.json>
python scripts/retrieval_contract.py validate-pair <task.json> <result.json>