| name | inbox-triage |
| description | Morning email triage — type /inbox-triage to classify the last 24h of Gmail into leads / urgent / warm / partners / meetings / noise, research real leads, draft (never send) replies, and drop a one-glance brief into the vault. |
| argument-hint | [hours to look back, default 24] |
| disable-model-invocation | true |
| allowed-tools | Read, Grep, Glob, Bash, Write, Agent, WebSearch, WebFetch, ToolSearch |
inbox-triage — bucket, research, draft, brief
Turns the morning inbox read into a five-minute glance: every message bucketed, real leads
pre-researched with a go/no-go recommendation, replies drafted for approval, one report in
the vault. Phill stays the arbiter on every send.
When to invoke
- Start of the working day, or after any >24h away from the inbox.
- Before a discovery call, to pull the sender's thread history and background in one pass.
- You are about to hand Phill a raw inbox summary with no recommendations — stop, run this.
Core procedure
- Fetch. Read the lookback window (default 24h) from Gmail. Route: check
library/connections.md for the live Gmail plane (claude.ai connector vs Composio
substrate — prefer Composio when both are up, per estate memory). If neither
authenticates, STOP and report the auth gap — never fabricate inbox state.
- Done when: message list with sender/subject/body-preview is in hand.
- Bucket. Classify each message: leads (new business for RA / CARSI / CCW / Unite /
Synthex), urgent (deadline, outage, founder-named), warm (ongoing client/partner
threads), partners (sponsorship/collab/vendor offers), meetings (invites,
reschedules), noise (newsletters, notifications, cold spam).
- Done when: every message has exactly one bucket.
- Act per bucket.
- leads → validate coherence (budget, timeline, real intent vs form spam); for real
ones run a background WebSearch on the company; output go/no-go + one-line why + a
drafted reply with the booking link.
- urgent → drafted reply each.
- partners → drafted boilerplate response pointing at the relevant rate/media info.
- warm / meetings → one-line status each; flag anything stale per the
stale-action-items rule (verify current thread state before drafting).
- noise → count only, listed at the report tail.
- Done when: every non-noise message has an action artifact or a one-liner.
- Report. Write the brief (format below) to the vault and surface drafts for approval.
- Done when: report file exists and the reply names where every draft lives.
Output format
~/2nd Brain/2nd Brain/Briefs/YYYY-MM-DD-inbox.md — frontmatter type: brief,
created: YYYY-MM-DD; sections per bucket in the order above; each lead entry:
sender, company, ask, budget signal, research digest (2-3 lines), GO / NO-GO + why,
draft location. Drafts are created as Gmail drafts when the connected plane supports it,
otherwise inline in the report under the entry. (Briefs/ is the vault's dated
agent-briefs folder — shared with daily-intel-brief; create it on first run.)
Calibration
- Full run ≤ 10 minutes, ≤ 3 web searches per lead. A lead needing more research than that
is a discovery-call question, not a triage question.
-
5 "urgent" items usually means over-triggering — urgent requires a deadline, an outage,
or the founder named in the ask.
What this skill is NOT
- Not
email-listener — that converts inbound mail into Pi-CEO sessions via webhook; this
is the on-demand/morning human-facing triage pass.
- Not a sender: it never sends anything. Draft + surface for Send is the decision-rights
boundary.
- Not
sow-draft/stripe-milestone-invoice — a GO lead hands off to those, it doesn't
absorb them.
Hard rules
- NEVER send. Drafts only. (Outbound is founder-gated estate-wide.)
- Never fabricate inbox state — if the mail plane fails, the report says so and stops.
(No-false-recordings directive 2026-07-15.)
- Verify thread state before drafting on warm items — action items observed to go stale.
- GO/NO-GO always carries the one-line why; a recommendation-free list is a violation.
Provenance
