feat(skills): add inbox-priority-agent triage skill

skills/inbox-priority-agent/SKILL.md

@@ -0,0 +1,62 @@
+---
+name: inbox-priority-agent
+description: Triage inbound messages and produce a priority-ordered action list with concise draft replies. Use when the user asks to review inboxes, mentions, DMs, email summaries, or mixed inbound queues; when they want urgency/importance sorting; or when they want low-noise recommendations in a Now/Soon/Watch format without sending anything automatically.
+---
+
+# Inbox Priority Agent
+
+## Overview
+Prioritize inbound communication into a short action queue, then suggest clear next moves and optional draft responses. Optimize for signal over noise and avoid unnecessary pings.
+
+## Workflow
+
+1. Collect context
+   - Identify source(s): chat, email, mentions, notifications, or mixed inbox.
+   - Capture sender, timestamp, thread context, and explicit deadlines.
+
+2. Classify each item
+   - Score urgency and importance using `references/priority-rubric.md`.
+   - Assign one level: `P1` (critical), `P2` (important), `P3` (normal), `P4` (low/no action).
+   - Mark confidence as `high`, `medium`, or `low` when context is incomplete.
+
+3. Deduplicate and compress
+   - Merge duplicate notifications and same-topic follow-ups.
+   - Suppress low-value repeats unless status changed.
+
+4. Build briefing output
+   - Present sections in this order:
+     - `Now` (P1)
+     - `Soon` (P2)
+     - `Watch` (P3/P4 changes worth awareness)
+     - `Next actions` (numbered, minimal, concrete)
+   - Keep each bullet to one short line.
+
+5. Draft replies only when useful
+   - Include drafts for P1/P2 items or when explicitly requested.
+   - Keep drafts short, polite, and directly actionable.
+   - Never send external messages automatically unless explicitly instructed.
+
+## Output Template
+Use this exact structure unless the user requests another format.
+
+- **Now**
+  - [item] — why it matters + deadline/risk
+- **Soon**
+  - [item] — why it matters + recommended window
+- **Watch**
+  - [item] — context only / no immediate action
+- **Next actions**
+  1. [concrete action]
+  2. [concrete action]
+
+If drafts are included, append:
+
+- **Draft replies**
+  - To [name/channel]: "[draft]"
+
+## Guardrails
+- Do not fabricate missing context; label assumptions clearly.
+- Prefer fewer, higher-signal recommendations over exhaustive dumps.
+- If uncertain between two priorities, choose the higher one and note uncertainty.
+- Treat private/sensitive data minimally and avoid unnecessary quoting.
+- Ask before any state-changing external action (send/reply/post).