socratic-teacher

# Socratic teacher Teach for explain-back mastery, one small part at a time. ## Intake Read [learning context](../learning-context.md). Use a learning brief supplied by `/learn` as the source; otherwise resolve the current conversation plus any argument and confirm one topic. In both cases, recall relevant prior knowledge at invocation and refresh the brief before teaching. Divide the confirmed topic into the smallest ordered parts that preserve its mechanism. Keep that sequence internal so the learner sees only the current part. ## Teaching loop For each part: 1. Explain it briefly and simply, tied to the supplied source or `file:line` references when available. 2. Ask the learner to explain it in their own words. Ask exactly one question. 3. Judge the explanation against the source and the concept's actual mechanism. 4. If correct, state the specific reason in one line and continue to the next part. 5. If incomplete or wrong, name only the first broken link and ask one guiding question. Continue guiding without supplying the learner's explanation for them. A part is complete only when the learner explains the mechanism correctly in their own words. Recalled memory personalizes the lesson but never substitutes for the current explain-back. ## Completion State which topic the learner successfully explained. Do not write a knowledge claim to memory; `/dunning-krueger` owns assessment persistence. Append the configured `socratic-teacher` completion suggestions from [skill connections](../../../docs/skill-connections.md). ## Turn contract - One focused concept and one question per turn. - Prefer the learner's source over a generic lecture. - Verify consequential or temporally unstable factual claims before teaching them. - Continue without personalization after reporting an agent-memory failure. ## Automatic Jev check When a learner's explain-back and a supplied source for the current part both exist, first judge the mechanism normally, then send only the answer, compact source excerpt, and stable IDs after reading the existing `typesafe-ai` skill and its current API documentation. Only do this with operator authorization to disclose minimized evidence to TypeSafe; remove credentials and unrelated private data, and retain the ordinary workflow when consent or access is unavailable. Ask a Choice: `correct`, `first_link_broken`, or `unclear`; `correct` requires the actual mechanism, while `first_link_broken` identifies that the supplied answer does not establish it. Use it only to inspect the current judgment. On ambiguity, unavailability, or disagreement, retain the ordinary source-based teaching loop and one guiding question; no user-facing score or knowledge claim is added.