Produce the weekly operations readout for the service desk: what came in, what got
closed, how clients felt, what is aging, and what looked unusual — a fast stats pass
first, then targeted deep dives only where something moved.
This runs MANUALLY on demand — Thread Flows are ticket-event triggered with no
schedule/cron, so this is not an unattended/scheduled skill; run it on request or from
an external scheduler that invokes Super Magic.
1. Confirm the period (default: last full week) and the boards in scope.
Exclude junk/NOC and internal-noise boards up front.
2. FAST STATS PASS — run split searches per signal per board (opened, closed, still
open, aged past threshold, negative sentiment) rather than one giant query. Record
which searches hit a result cap; never present a capped result as an exact total —
disclose caps and give order-of-magnitude estimates instead.
3. Report headline numbers vs the prior week: volume in, closures, net queue change,
aging count, sentiment trend. Exclude auto-closed statuses from closure credit so
automation noise doesn't inflate the numbers. Do not fabricate a prior-week
baseline; if last week's data is unavailable, report this week standalone and say so.
4. ANOMALY SCAN — flag anything that moved sharply week over week: a volume spike on
one board or client, a closure-rate drop, a sentiment dip, a growing aging bucket.
5. TARGETED DEEP DIVES — for each flagged anomaly only, pull a narrow follow-up search
and explain the likely driver in one or two sentences (alert storm, one large client
incident, a tech out sick per the queue pattern). AGGREGATE rather than enumerate —
patterns and counts, not ticket lists; name a specific ticket only when it explains
an anomaly. Per-tech numbers are context, not rankings — benchmark role-aware and
never flag a person on volume alone.
6. End with 2-3 recommended actions and a methodology note: period, boards
included/excluded, searches run, and any result caps hit. Use fixed sections
(Headlines / Anomalies / Actions / Methodology). If a search fails or returns
nothing, state "no data for <section>" rather than guessing.
