meeting-classifier

Classify a teacher's meeting-related request, then route it. Use whenever a teacher mentions a meeting or invite — faculty, grade-level/PLC, department, data, or emergency meetings; evaluation observa

SKILL.md

GitHub

meeting-classifier

Classifies a teacher's meeting + what they want done about it, then routes to the right TOS skill — decision support, not a determination. It classifies + routes (and produces a light prep brief); it does not author the meeting artifacts (sibling skills do).

Read first — boundaries (SECURITY_AND_SAFETY.md). This skill never makes a determination about an individual student's IEP/504, eligibility, medical, or legal status — it escalates those to a human/team. Medical/ePHI is surfaced from the source on file (attributed; a signature is not required), never fabricated (references/student-data-policy.md); real student data never enters a tracked/committed file.

1. Resolve context, connectors, and student first

  • Context — resolve the teaching-context contract (shared/context/), including the department scope; carry it through the handoff.
  • Connectors — read what's available (shared/connectors/): use whatever is connected, degrade to what you have permission for, and converge via alternate sources; never present a disabled connector as an active path; lower confidence on degraded paths.
  • Student — when a meeting concerns a student, resolve the profile (shared/students/, identification mode name-by-default; SIS authoritative when connected). Surface guardian contacts for a call home, and a medical action plan from the source on file when the meeting/call is medical (safety-critical).

2. Follow the pipeline (references/method.md)

Within Generation the domain work is classification:

  1. Gather clues from the available evidence per references/evidence-model.md (strong→weak: explicit statement, calendar event, sender title/role, attendee roles, subject/body, prior thread, transcript/notes, filename).
  2. Classify the meeting type (one of the taxonomy in references/meeting-taxonomy.md).
  3. Classify the request intent (prep, draft communication, summarize/minutes, identify attendees, schedule/triage, documentation/compliance, extract-only, or unknown).
  4. Attach advisories — for iep_meeting/section_504_meeting attach the required-cadence advisory (an existing plan is reviewed on a legal cadence; verify_on_source; cite IDEA / Section 504); for a medical meeting, surface the action plan from the source on file and flag safety-critical.
  5. Decide honestly — when two meeting types are materially plausible (e.g., annual-review vs interim observation) or sources conflict, emit a minority report via the resolver (shared/context/sot_resolver.py; policy shared/context/minority-report.md) instead of guessing; when evidence is too weak, return unknown and recommend manual_review / one clarifying question.

3. Route (don't author)

Use the routing table in references/meeting-taxonomy.md (parent contact → family-communication; IEP/504 → special-education-support + escalate; MTSS → intervention-mtss; observations → professional-learning or school-administration; planning → curriculum-mapping/lesson-planner). Only recommend a skill that exists, or manual_review. The bundled scripts/classify_meeting.py gives a fast first-pass label from compact clues; if it conflicts with stronger evidence, follow the evidence.

4. Artifacts

See references/artifact-types.md — a meeting classification record, a routing recommendation, and a light prep brief / handoff packet (with guardian-contact info + a medical safety banner when relevant). No full artifact authoring here.

5. Output: always emit the metadata block

End with the metadata block from references/metadata-schema.md, including the quality decision and human_review_required: true — outputs are decision support, not final professional, medical, or legal determinations. Placeholders only in any committed example; real student PII/ePHI stays in the storage adapter (shared/students/), never committed.