Email triage
The shared-inbox triage runbook — how to read a new inbound email, label it, look up {{help_doc}} for a standard answer, draft a reply in Gmail or file a ticket in {{linear_team}}, and when to leave a thread label-only for a human.
Overview
Every inbound email in the shared inbox gets read, labelled, and routed to one of three outcomes: a drafted reply, a filed Linear ticket, or a label-only hold for a human. A scheduled check spawns a fresh session with scoped, read/write access to Gmail and Linear and read access to {{help_doc}}; this skill is the standard the triage is done to, so the agent doesn't reinvent categories or tone on every run. A single check can turn up several new messages at once — handle each as an independent unit: a failure or bad match on one message never blocks triage of the others found in the same pass. Nothing customer-facing ever sends from this skill — a draft is the most it produces.
When to load
- The scheduled inbox check fires and there is new inbound mail to triage.
- A human asks why a specific thread got the label, draft, or ticket it did.
- A human asks the agent to re-triage the current inbox on demand.
Workflow
Step 1 — Pull what's new
Check the shared Gmail inbox for messages that arrived since the last check and don't already carry a triage label. Read each thread in full, not just the latest message — a reply-in-progress needs the whole context.
Step 2 — Categorize the message
Classify every thread into exactly one category:
| Category | Signal | Label |
|---|---|---|
| Support question | Asking how something works, a how-to, an account question with a known answer | triage/support |
| Bug report | Something broken, an error, unexpected behavior | triage/bug |
| Sales lead | Pricing, demo, or new-customer interest | triage/sales |
| Billing / invoice | Payment, invoice, or subscription question | triage/billing |
| Other / spam | Not a real request, or nothing the team needs to act on | triage/other |
Apply the label in Gmail regardless of what happens next — every message that's triaged gets exactly one category label.
Step 3 — Look up the help doc for support questions
For triage/support, search {{help_doc}} for the closest matching answer before writing anything. Match on the actual question asked, not just keywords — a doc section that's adjacent but doesn't answer the specific question isn't a match.
| Match quality | Action |
|---|---|
| Doc directly answers the question | Draft a reply (Step 4) |
| Doc partially answers it, or the question combines two topics | Draft a reply covering what the doc supports; note in the draft what's uncertain |
| No relevant section in the doc | Label only — leave it for a human, don't guess |
Step 4 — Draft the reply (never send)
Write the reply in the sender's language and tone: direct, helpful, answer first. Cite or paraphrase the help-doc section it's drawn from so the human reviewer can verify it quickly. Save it as a Gmail draft on the thread — this is the only thing that touches the customer's view of the thread, and it is not sent until a person sends it.
Step 5 — File bug reports and follow-ups in Linear
For triage/bug and any triage/support thread that needs engineering follow-up rather than a reply, create a ticket in the {{linear_team}} team:
- Title: a short, specific summary of the reported problem.
- Description: the sender, the thread's key details, and a link back to the Gmail thread.
- Priority: bump to urgent only if the sender describes a full outage or data loss; default to normal otherwise and let the team triage.
Sales leads (triage/sales) and billing questions (triage/billing) get labelled and, if the help doc covers the billing question, drafted per Step 3 — they don't automatically get a Linear ticket unless they also describe a bug.
Step 6 — Leave triage/other alone
Label and move on. No draft, no ticket, no further action.
Step 7 — Stop
One pass over the new mail is one run. There's no ledger to update — the next scheduled check re-reads whatever the inbox looks like then.
Guardrails
- Draft only, never send. Every reply is a Gmail draft on the thread. The agent never sends a customer-facing email itself, no matter how confident the match to {{help_doc}} is.
- No memory between runs. Each check is a fresh session; the current inbox and label state is the only source of truth, not what a prior run concluded.
- Don't guess past the doc. If {{help_doc}} doesn't clearly answer a support question, label it and stop — a wrong drafted answer waiting for rubber-stamp approval is worse than no draft.
- Scoped secrets. Gmail, Linear, and help-doc access are brokered through connectors; no raw token is ever shown to the model or written to logs.
- People decide, not the agent. The agent labels, drafts, and files — whether a draft sends or a ticket's priority changes is a human call.
