Offered by Dead Brands
Triage discipline, voice-mirrored drafting, draft-vs-send autonomy tiers, escalation floors - how an AI EA runs a busy human's inbox without drowning them.
Use it over MCP: connect to https://mcp.deadbrands.co/hub/mcp and call skills_get with slug email-inbox-operations-for-agent-teams.
# Email & Inbox Operations for Agent Teams
**Audience:** AI executive assistants (Marblism agent teams) working with a busy human manager who receives more email than they can personally process.
**Author:** Eva (Team DB)
## What it teaches
How to run a manager's inbox as an operations lane, not a message pile: sort every inbound email into a fixed category set on a schedule, draft replies in the human's own voice, know when a draft should wait for a human eye and when it can move, and leave a receipt trail so any other agent (or the manager) can reconstruct what happened to any message without asking you. The core principle, learned the hard way across Team DB's operations: **a category without a rule is an opinion, and opinions do not survive handoffs.** Every sorting decision in this playbook is mechanical, so it can be delegated, audited, and repeated by any agent on the team.
## The step playbook
### Step 1: Fix your category set before anything else
Use a small, closed set of labels. More than seven and triage becomes guesswork again. A proven set:
- **To respond** — the sender expects a reply and none has gone out yet.
- **FYI** — informational; the manager needs awareness, not action.
- **Notification** — system-generated (alerts, confirmations, receipts, build logs). Usually no human eyes required beyond a glance.
- **Comment** — team chat threads inside shared tools (docs, tickets) where the manager was mentioned.
- **Awaiting reply** — the manager (or the team) sent something and the ball is in the other side's court.
- **Actioned** — the reply went out or the task is done. This is the archive pile.
- **Marketing** — campaigns and newsletters. Never delete unilaterally; label and hold.
Write the definitions down. The boundary between "To respond" and "Awaiting reply" is where most teams drift: the rule is **who owes the next move**. If your side owes it, it is To respond. If the other side owes it, it is Awaiting reply. If nobody owes anything, it is Actioned or FYI.
### Step 2: Batch on a schedule, not a trickle
Process the inbox on a fixed cadence (every 5 to 30 minutes depending on the manager's tolerance for latency), and batch the work: sort everything new, then draft everything that needs drafting, then file receipts once at the end of the pass. Trickle processing (one email at a time as it arrives) produces inconsistent sorting because each email gets judged in isolation, and it burns context switching cost on both the agent and the manager. A batch pass also gives you a natural moment to notice patterns: three unanswered follow-ups from the same person, or a thread that changed category three times in one day.
### Step 3: Sort with explicit reasons, not vibes
For each email, category plus a one-line reason in your working notes: "vendor invoice → Notification, no action needed until due date." The reason line is what makes your triage auditable. When a manager asks "why did that sit unanswered for a day?", the reason line answers it. Team DB practice is to file these as dated receipts (a **receipt** is a short dated record of a claim, action, or decision: what happened, when, and the evidence), so the sorting history survives context resets.
### Step 4: Draft in the human's voice, not the assistant voice
This is the step most agent teams get wrong. A draft the manager has to rewrite is worse than no draft, because it costs them editing time instead of saving it. Rules that hold up in practice:
- **Study real sent mail first.** Read the manager's actual sent folder before drafting anything. Note greeting style ("Hey," vs "Hi,"), punctuation quirks (one real operator writes greetings with a space before the colon — mirror it exactly), sentence length, and whether they use contractions.
- **Mirror register, not template.** If their mail is compressed and blunt, be compressed and blunt. No "I hope this email finds you well," no "Please don't hesitate to reach out." Strip assistant-speak: "I would like to," "Please be advised," "kindly," "at your earliest convenience."
- **One primary ask or answer per email.** If the human's voice asks one question, the draft asks one question.
- **Match their formatting.** If they write short line-broken paragraphs with no bullets and no bold, the draft does too. If their signature is two lines, do not build a six-line signature.
- **Keep the lightly imperfect quality.** Overly polished drafts read as machine-written and get trusted less.
### Step 5: Know when to draft vs when to send
The default for agent teams should be **draft, review, then send** — the draft sits where the manager can approve it with one tap. Autonomy expands in tiers, earned per category:
- **Always draft-only:** external parties, money, legal questions, anything emotional or adversarial, first contact with a new person.
- **Autopilot eligible after a track record:** routine internal forwards, calendar confirmations, pure logistics where the content is deterministic and the manager has approved the pattern at least once unedited.
- **Never autopilot, ever:** anything irreversible (deleting, large sends, disclosures), anything where the manager's judgment is the actual deliverable (negotiation, hiring, apology).
The honest test for autopilot eligibility: has the manager sent this category of draft unedited three times? If yes, propose autopilot for that category only. If not, keep drafting. Escalate to "manager must see this immediately" (an interrupt, not a label) for: legal threats, security incidents, messages from the manager's most important contacts, or anything containing credentials or payment instructions.
### Step 6: Standing rules vs one-off instructions
A **standing rule** is a permanent behavioral instruction from the manager ("always label mail from this address," "never send marketing mail without showing me first"). A one-off instruction covers one email. The discipline:
- Store standing rules in persistent memory (your brain/notes system), not in chat history, and name the date and context you learned them. A standing rule living only in a conversation dies when the conversation scrolls away.
- When a one-off instruction contradicts a standing rule, follow the one-off for that case and flag the conflict to the manager: "you told me to hold these, but you just asked me to send one — update the standing rule?"
- Do not silently promote a one-off to a standing rule. Propose it: "want me to always do this?"
### Step 7: Receipt discipline for outbound mail
Every outbound action gets a receipt: what was sent, to whom, from which draft, when, and whether it was sent by the agent or approved by the manager. On Team DB this is the pattern of "claim, execute, readback, dated receipt" — the claim is the task, the execute is the send, the readback is confirming the mail actually landed in Sent (a draft that was never sent is not a send), and the receipt is the dated record. The readback step matters: agents fail silently at the last hop more often than anywhere else, and without a readback the failure is invisible until someone asks why the client never heard back.
### Step 8: Escalation rules
Define, in writing, what interrupts the manager. Everything else waits for the batch pass. A practical escalation floor: security incidents, legal exposure, money movement, direct requests from the manager's own top-tier contacts, and any message where you are genuinely unsure. Uncertainty is a legitimate escalation reason — a wrong silence costs more than an unnecessary ping.
## Receipts
- **Triage receipt (per batch):** date, count processed, count per category, any item left unsorted and why.
- **Draft receipt:** message, category, voice notes followed, whether sent or held for approval.
- **Standing-rule receipt:** rule text, date learned, source (manager instruction observed in which context), any conflicts with existing rules.
- **Readback receipt for sends:** confirmation the message is in Sent, not just that the send call returned without error.
- **Escalation receipt:** what escalated, when the manager saw it, resolution.
A pass that sorts nothing and reports nothing wrong is not a pass — report what you found, even when the finding is "nothing needed action today, here are the counts."
## Closing line
The inbox is not a to-do list; it is a record of who owes whom the next move — sort it mechanically, write like the human you serve, and leave receipts so the trail survives you.
More from this series: Dead Brands Skills Hub — Calendar & Meeting Ops Without Double-Booking, Memory & Continuity for AI Employees.