Calendar Defrag Skill

Calendars fragment the way disks did: nothing big fits anywhere, and every week costs more than it returns. This skill audits the week like a capacity planner (what each meeting costs, what the fragments waste), proposes the defrag, and — only with approval — executes the moves through an agent with calendar access.

What This Skill Produces

• The audit — every recurring meeting priced (hours × attendees), fragmentation map of the week, the largest contiguous focus block currently possible
• The defrag proposal — meetings to shorten / merge / make async / decline-with-note, moves that consolidate fragments into 2h+ focus blocks
• The execution plan — exact calendar operations, shown before any are made

Required Inputs

Ask for these if not provided:

• Calendar scope — which week(s), work hours, timezone
• The user's real priorities — defrag serves deep work on something; name it
• Untouchables — meetings that are politically or contractually fixed
• Meeting-owner etiquette — may the agent propose times to others, or only move solo/owned events?

Framework

1. Price everything: duration × attendees = cost; recurring cost × 26 = the six-month bill. Print it — the bill changes minds that advice can't.
2. The fragment rule: a gap under 45 minutes is lost time; count the week's fragment waste in hours.
3. The defrag moves, in preference order: delete (with a note) → make async (doc + comments) → halve the duration → merge with an adjacent ritual → move to the edge of the day.
4. Focus blocks are placed first — 2h+ blocks at the user's stated peak hours; meetings pack around them, not vice versa.
5. Respect other people's calendars: moves affecting others become proposals (draft messages), never unilateral moves.

Output Format

Calendar Defrag: week of [date]

The bill: [n] meeting-hours (× attendees = [n] person-hours) · fragments waste [n]h · largest focus block today: [n]min

Meeting	Cost	Verdict	Move
After defrag: [n] focus blocks ([hours]), [n]h returned per week.

Quality Checks

• Every recurring meeting carries its six-month price
• Every removal has a draft note to the organizer — silent declines burn trust
• Focus blocks land on stated peak hours, not leftovers
• Moves touching other people are proposals with drafts, not actions

Anti-Patterns

• Do not judge meetings by name — a "sync" can be load-bearing; price and ask what it decides
• Do not create focus blocks and leave them unlabeled — an empty slot gets colonized in a day
• Do not defrag someone into back-to-back meetings to serve the user — fragmentation exported is not fragmentation solved
• Do not touch the untouchables list, even when they're obviously the problem — name it, don't move it

Execution

For agents with calendar access (Google/Microsoft APIs or UI). Without tools, the audit + proposal is the deliverable. Rules per SKILLSPEC.md §5.

Preconditions

• The defrag proposal has been produced and explicitly approved by a human, line-item vetoes honored.
• Calendar access authenticated; the week(s) in scope named; the untouchables list confirmed.
• The execution plan (exact create/move/shorten/decline operations) displayed and confirmed.

Allowed actions

• Create the approved focus blocks (named, with decline-meetings visibility set as approved).
• Move or shorten only events the user owns, per the approved plan.
• Decline events on the approved list, attaching the approved note.
• Save (never send) the drafted proposals for meetings owned by others.
• Nothing else: no deleting events outright, no touching other calendars, no auto-accepting, no changes outside the approved week(s).

Verification

• Re-read the week: focus blocks exist at approved times; moved events at new times; declined events show the note; nothing outside the plan changed.
• Report the before/after fragment-waste numbers.

Rollback

• Every move is recorded (event, old time, new time) — undo restores old times and removes created blocks.
• Stop and ask a human if: an event was modified by someone else after approval, a move conflicts with a new event, or any API action fails.