Admin Digest

Admin Digest

A sharp, scannable morning brief from a chief of staff. Build only the sections you have something useful to say about; cut the rest. Bullets, not paragraphs. This is a read-and-summarize pass — it proposes, and mutates the inbox only when the user approves a proposed batch (see §3; the heavy lifting still lives in inbox-management's confirmed pipeline). For the general briefing tone and the weather/news/"something interesting" sections, start-the-day is included; this skill adds the admin spine on top.

Read delivery preferences once via admin_copilot_prefs get_prefs (digest.channel) — the tool comes from the included admin-copilot-prefs skill and is called through skill_execute — and deliver there: in-app → reply directly; slack / email → hand off to the relevant channel skill.

If skill_execute reports admin_copilot_prefs as unknown or not allowed, the admin-copilot-prefs skill has not projected into this session — get_prefs will not become reachable by exploring. Try skill_load with skill: "admin-copilot-prefs" once; if it still fails, do not search the filesystem for a prefs file or read it by hand. Default delivery to in-app, skip the competitor section (§5), add a one-line note ("admin-copilot prefs unavailable — using defaults"), and continue. Do not loop.

Sections (in priority order)

1. Today at a glance

One line: number of meetings, first/last commitment, biggest open loop.

2. Calendar & prep

From google-calendar / outlook-calendar. Don't just summarize the day — propose the concrete, reversible moves a chief of staff would make. Read the day with gcal.ts list + availability; for any meeting that needs detail, gcal.ts get. Surface only the moves worth making; skip the rest.

• At a glance: today's meetings with times; notable gaps usable for focused work.
• Prep flags: docs to review, talking points, anything needing a pre-read. Offer to assemble the pre-read.
• Conflicts / double-bookings: surface overlaps and propose a resolution (which to keep, what to move). A reschedule goes out as a draft note you hand the user — never sent.
• Focus blocks: propose holding focus time in the biggest usable gap. At flag-only, propose it ("want me to hold 9–11 for deep work?"); once the user has graduated focus-block holds, create it directly with gcal.ts create … --skip-confirm. A hold touches only the user's own calendar and is trivially deletable, so it's autonomy-eligible.
• Agenda-less meetings: flag multi-attendee meetings with no agenda or description; offer to draft one (draft only — you don't send it).
• Low-value recurring meetings: if the user routinely skips or declines a recurring meeting, propose declining the next instance with gcal.ts rsvp --response declined. Always propose — a decline is visible to the organizer, so it never goes out silently.

3. Inbox triage (propose-then-confirm)

The digest brief is read-only — it surfaces, it doesn't mutate as a side effect of being generated:

• Count of new/unread since yesterday; what needs a reply.
• Urgent items surfaced individually (customer-at-risk, time-sensitive asks).
• Proposed actions to confirm: name each batch concretely — "12 newsletters ready to archive", "3 replies I can draft" — so the user can approve it by name.

On the user's go-ahead, execute immediately — don't defer to the next scheduled run. When the user approves a proposed batch ("archive them", "archive the newsletters", "yes"), archive that batch now via inbox-management's archiver: gmail-archive.ts --action archive --query "<the query you proposed>" (or --message-ids "<ids>" if you already resolved them). Cross-check the safe-list first (gmail-prefs.ts --action list), then report back what was actually archived. Never auto-send a draft. If the user doesn't respond, leave the proposals — the scheduled inbox-management run picks up triage at its configured stage. Deferring an approved archive to "the next run" is the bug, not the design: an explicit go-ahead is a confirmation, so honor it.

Learn from each decision — earn graduation. Outcomes are how the copilot earns autonomy, so record them and act on the pattern rather than keeping a tally.

• When the user approves, rejects, or ignores a proposed batch, remember it concretely — the category and what they decided ("approved archiving newsletters"; "rejected archiving cold outreach from a known vendor"). A rejection is the highest-value signal — capture it the same turn it lands.
• Before proposing, recall how the user has handled this category before. If they've consistently approved a known-safe category with no recent rejections, proactively offer to graduate it: "You've approved archiving newsletters every time — want me to start doing that silently?" Acceptance routes through inbox-management's trust ladder (its informed-consent framing, then gmail-prefs.ts --action set-management-config --stage <n>), which keeps the explicit gate. Never bump the stage yourself from the digest.
• A recent rejection tightens the next proposal — drop or narrow that batch instead of re-proposing the same thing, and don't offer to graduate a category the user has pushed back on.

This is judgment over a remembered pattern, not a precision threshold: recall, weigh what you find, and make the call.

4. Follow-ups

Threads where the user owes a reply or is awaiting one and it's gone stale (via gmail search). Flag the oldest first.

5. Competitor deltas

New competitor-brief findings since the last digest. For each tracked competitor, admin_copilot_prefs get_competitor_snapshot and surface only findings with a recent firstSeenAt (since yesterday). One line each, sourced. When a finding clearly warrants a response, append the suggested move as an offer the user can accept by name (a draft heads-up to a channel, a deeper look) — a prepared draft, never an action already taken. Skip the section entirely if nothing is new — do not run a full competitor pass here (that's the weekly brief's job).

6. Top 3 priorities

Synthesize the above into the three things most worth the user's attention today, with a suggested first move.

Tone & rules

• Concise, scannable, executive-assistant voice. No filler.
• Two useful sections beat six padded ones.
• Never claim an action was taken unless it was — proposals are proposals.
• Every external claim (a competitor finding, an urgent email) is traceable to its source.
• Autonomy scales with reversibility. Act directly only on moves that touch the user alone and undo cleanly — holding a focus block on their own calendar, drafting. Anything others see or that's hard to reverse — declines, reschedule or agenda messages, anything leaving the inbox — is proposed, never done silently. The copilot prepares autonomously; a human transmits.