icaire-learn-reply-to-forum-posts

Retrieve ICAIRE Educate forum posts, draft a batch of distinct instructor replies for approval, and send only exact approved replies through the OAuth-connected ICAIRE Educate MCP. Use when the user a

SKILL.md

GitHub

ICAIRE Learn: Reply to Forum Posts

Prepare and send thoughtful instructor replies to ICAIRE Educate discussion threads with a hard human approval gate before every write.

Contract Checklist

  • Use the OAuth-connected ICAIRE Educate MCP as the source of truth so the signed-in instructor identity determines the author; never choose or expose raw admin identifiers when OAuth can resolve identity.
  • Expect list_cohorts, list_cohort_forum_posts, and reply_to_cohort_forum_post; use find-missing-tools when the tool surface is incomplete.
  • Default to at most 10 unanswered top-level threads, oldest first.
  • Exclude top-level posts authored by an instructor, administrator, or the OAuth-resolved active identity from the default reply queue.
  • Keep learner-authored threads in the queue when peers have replied but no instructor or administrator has replied.
  • Draft genuinely varied responses rather than copying repeated wording verbatim across learners.
  • When explaining where to find something, include a Markdown link to the relevant course section.
  • Show the exact reply batch and obtain explicit approval before sending.
  • Re-fetch before sending and read back every successful reply afterward.
  • Use post_cohort_forum_message only when explicitly asked to start a new discussion; never substitute it for a reply.

Workflow

  1. Resolve the forum queue:
    • resolve the requested cohort; if none was named and several are relevant, show the choices and ask which one to use
    • call list_cohort_forum_posts with unanswered_only: true by default
    • use a supported server-side author filter when available to exclude top-level posts by instructors, administrators, or the OAuth-resolved active identity; otherwise filter them locally before loading full detail
    • treat staff-authored top-level posts as prompts to monitor, not learner messages needing instructor replies; include them only when the user asks to review staff prompts, engagement, moderation, or replies to them
    • preserve learner-authored threads when learner peers have replied; for this queue, a thread is answered only after an instructor or administrator response, not merely because any reply exists
    • determine staff authorship from explicit author role and OAuth-resolved identity, never from writing style or a guessed name
    • select at most 10 top-level threads and sort them oldest first
    • include answered threads only when requested and label them as follow-ups
    • keep exact post IDs internally, but show stable labels such as R1, R2
    • Anti-patterns: mixing cohorts without consent, queueing our own admin prompt, treating a peer reply as an instructor answer, guessing identity, loading full staff threads only to discard them, replying to nested IDs, silently including answered threads, bypassing the MCP
  2. Draft distinct replies:
    • read each full thread, including replies, author role, timestamps, and useful vote context
    • answer the learner's actual question in a concise, warm instructor voice
    • acknowledge relevant reasoning, distinguish facts from interpretation, and give a practical next step where useful
    • vary openings, sentence structure, acknowledgements, and closings when several learners ask similar questions; do not paste one answer verbatim
    • use supported Markdown links such as [section name](https://...) when a learner asks where to find course material, a quiz, or another resource
    • flag missing facts instead of inventing policy, dates, grades, certificates, attendance, commitments, partnerships, or technical claims
    • for private support or safeguarding matters, draft a brief public acknowledgement and mark the required private follow-up
    • Anti-patterns: templated duplicate replies, bare URLs when a descriptive course link is available, invented facts, exposing private learner data
  3. Present the batch for approval:
    • for each item show its label, cohort, thread excerpt, learner's point, proposed reply in full, confidence or missing context, and action
    • state clearly that nothing has been sent
    • accept explicit choices such as approve all, approve R1 and R3, edit R2: ..., skip R4, or stop
    • after a material edit, display the revised wording and require approval again; approval applies only to the displayed wording and mapped thread
    • Anti-patterns: treating praise as send approval, hiding full wording, carrying approval from another batch, editing silently after approval
  4. Send only approved replies:
    • re-fetch each approved thread immediately before sending and confirm the top-level post still exists, matches the reviewed content, and has not gained an instructor response
    • rely on OAuth identity resolution; if the MCP cannot unambiguously resolve the signed-in instructor, stop before the first write
    • call reply_to_cohort_forum_post once for each approved, current thread
    • continue safely after an individual failure without retrying successful items or sending skipped items
    • re-list the threads and verify each successful reply under the intended top-level post with the resolved instructor identity
    • Anti-patterns: sending stale drafts, duplicate retries, manual identity substitution, direct database writes, browser form submission

Anti-Patterns

  • Sending anything before explicit approval of the exact displayed draft.
  • Queueing an instructor's or administrator's own discussion prompt as a learner message needing a reply.
  • Removing a learner thread solely because another learner replied.
  • Replying to a nested reply ID or starting a new discussion instead.
  • Reusing identical prose across multiple learners.
  • Claiming success without MCP read-back verification.
  • Exposing learner data, access tokens, raw MCP responses, or admin IDs.

Output

Before approval, report the cohort and learner-only queue scope, numbered full drafts, missing context or private follow-ups, confirmation that nothing was sent, and the available approval choices.

After sending, report sent labels and thread summaries, skipped or stale items, per-item failures, MCP read-back status, and any private follow-up still needed.