CASE STUDY · 02 OF 03
Six sources pulled in parallel — email, calendar, team chat, operations databases, the task tracker, and Granola meeting notes. Claude reads the day's signals against a written spec, a second pass surfaces repeating themes worth a second look, and the briefing renders inline before the founder finishes a first coffee. A dated session log is archived for the next day's lookback.
01 ·Context
The morning routine had stretched to nearly an hour: priority email scanned for time-sensitive senders, four calendars checked for the day's availability, two chat tools combed for DMs and watched channels in the last twenty-four hours, five operations databases opened to surface case-tracking activity, the task tracker queried for anything urgent or overdue, and Granola meeting notes reviewed for action items from the day before. By the time the first meeting started, half the morning was gone and the briefing was held in working memory.
The cost was not any single source — each was tractable on its own. The cost was the stitching: holding six threads at once, deciding what was a signal, deciding what to act on, all before the day had really begun. The same shape of work, every weekday, for years.
02 ·The system
“Run morning briefing.” That phrase, typed once, is the only input. A preflight step loads the runtime config — voice rules, priority-sender list, watched channels, the rolling lookback windows — before any source is touched. No schedule in v1; the founder pace is the schedule.
Email scanned by priority sender, time-sensitive keywords, and CC patterns. Four calendars checked for the day's availability. Team chat queried across watched channels and DMs over a twenty-four-hour window. Five case-tracking operations databases pulled for fresh activity. The task tracker queried for urgent and overdue items. Granola meeting notes pulled for action items from the day before. Every connector is read-only by design.
Claude reads the day's raw material against a written spec — what counts as a signal for this operator, in this voice. A second pass scans for repeating themes (three or more related signals in the last week) so a pattern doesn't hide behind individual rows. Two separate reads, one for what the day says, one for what the week is saying.
The briefing composes against the voice spec, runs a format pass, and renders inline in the same Claude session the founder triggered the run from. Nothing is mailed, Slack-posted, or sent on the founder's behalf — the inline render is the delivery surface. The briefing is read in place and acted on the same session.
Every run writes a dated session log — the assembled raw material, the classified output, the clusters surfaced, the briefing as rendered. The next morning's run reads yesterday's log as part of its lookback window so carryover and follow-through are visible without being re-derived. Read-only on sources; write-only on the log.
03 ·The build
Founder-operated, founder-built — the workflow that pays for the hour the founder used to spend stitching it together by hand. Phase by phase, what landed and what was learned.
04 ·Result
The morning routine collapsed from close to an hour of parallel tab-switching to a single phrase and an inline read. The classifier surfaces the signals; the cluster detector surfaces the patterns; the briefing renders in place. The founder reads, decides, acts. The first meeting is the next thing on the calendar, not the next thing after the morning routine.
Carryover is the underrated win. Yesterday's session log feeds today's preflight, so follow-through is visible without being re-derived. A theme that surfaced three days ago is still there if it's still relevant, and gone the moment it isn't.
05 ·Stack
Gmail·Google Calendar·Slack·Granola·Cowork·Claude·Skills·MCP