Find recurring patterns in the user's recent work signals that could be turned into leverage — a blog post (the user keeps explaining the same thing) or an automation (the user keeps doing the same manual, multi-step task). Give blog and automation candidates equal weight; bias toward automation candidates, since offloading repetitive work is usually the bigger win.
Default lookback window: the last 7 days from the current date/time. Treat the provided current datetime as authoritative — never reason from training-data dates.
Tool names. This skill refers to Microsoft 365 tools as
m365_*. If your host exposes them under different names, map them to the equivalent capability.
Inputs — collect breadth, not depth
- Email — list Inbox and Sent from the last 7 days (subjects + first-line previews only; do not open bodies). Pay special attention to Sent items: they reveal the user's own repeated outbound actions (forwards, intros, status pings, FYIs, "looping in X" handoffs, RSVPs, reschedules).
- Active chats — list recent chats; for each chat the user has posted in over the last ~30 days, fetch the last 7 days of messages. Strip @-mention noise.
- Active channels — for channels the user has posted in over the last ~30 days, fetch the last 7 days of messages. Skip silent channels.
- Calendar — list events from the last 7 days (subjects only — no bodies, no attendee lists). Look for repeated reschedules, repeated tentative responses, recurring conflict patterns, and meeting-prep cadences.
Clustering — what counts as a pattern
A signal qualifies only if it appears ≥3 times across ≥2 distinct people / chats / customers / threads / events in the window. Be ruthless — one-offs are noise.
Blog cluster types
- Repeated explanation — the user answers the same question or explains the same concept to 3+ different people. → blog candidate.
- Recurring FAQ — the same question shows up in 3+ chats/channels from different askers. → blog / FAQ candidate.
Automation cluster types (cast a wide net)
- Repeated multi-step manual task — the same sequence of clicks / drafts / file moves / API calls, repeatedly. → automation candidate.
- Repeated handoff / triage — forwarding, "looping in", or intro'ing the same kinds of items to the same people 3+ times. → auto-route by rule or classifier.
- Repeated status pings / FYIs — the same kind of update to the same audience on a cadence. → scheduled digest.
- Repeated meeting reschedules for the same root cause. → auto-propose alternates.
- Repeated RSVPs / tentative responses to recurring meetings with consistent reasoning. → auto-respond by rule.
- Repeated meeting prep — gathering the same context before recurring meetings. → pre-meeting brief.
- Repeated doc/wiki updates — manually appending the same kind of content on a cadence. → scheduled ingest.
- Repeated approvals / sign-offs of the same class of request. → rule-based pre-screen.
- Repeated file/data shuttling — move/rename/reformat between systems. → pipeline.
- Repeated drafting — near-identical emails/messages (welcome notes, onboarding intros, availability, "I have a conflict" replies, follow-ups). → template + auto-fill.
Anti-patterns — do NOT surface
- Normal recurring meeting cadence (standups, weekly syncs) UNLESS there's repetitive manual work around them that could be automated.
- Anything needing the body of an excluded/private email to reason about.
- Automations the user already runs.
Output
List 0–3 candidates max. If signal is thin (< 3 reinforcing data points
across all clusters), output: ✅ No new patterns this run.
Prefer a mix — include at least one automation candidate when one qualifies, even if a blog candidate is stronger.
Per candidate (≤3 lines each):
💡 **<theme in 5–8 words>**
_evidence:_ <1–2 anonymized signals, e.g. "forwarded 4 customer asks to the
same engineer this week" or "explained X to 4 different people in chats">
**<blog|automation>** · <1-line rationale. For automation, name the trigger
and the action, e.g. "trigger: incoming request → action: auto-route + draft
reply".>
Privacy & guardrails
- Read-only. Never send any email, chat, channel reply, or Teams message from this skill. Producing the report is the only output.
- Anonymize. Never include customer/team names, attendee names, meeting times/locations, file paths, or full subjects containing customer identifiers. Say "3 different customers", not the actual names.
- Respect exclusions. Never open or paraphrase content the user has marked as private or excluded (e.g. performance-review or confidential emails).
- When in doubt that a candidate is a real pattern vs. coincidence, drop it — silence is safer than noise.