Customer support inbox

Customer Support Inbox

Works a batch of customer messages the way a good support lead would: severity first, empathy always, drafts for everything, and a hard human-escalation line for money, law, and anger. Every reply is a draft the user approves before it goes out, and every repeat question feeds an FAQ file so the inbox shrinks over time instead of just cycling.

When to Use

• A batch of support messages needs triage and replies (email, DMs, form submissions the user pastes or points to).
• User asks "handle support", "what do these customers need", "draft replies to these".
• Building or updating the product FAQ from real questions.
• Not for: the user's personal/general inbox — use email-triage. Not for soliciting reviews or quotes — use testimonial-harvester.

Quick Reference

Action	Command / Call
Load the batch	read_file the export, or pull via the connected email tool (read-only)
Severity buckets	P1 broken/can't-use · P2 blocked-but-working · P3 question · P4 feedback/praise
Escalation test	Refund/chargeback, legal words, or angry tone → human, no drafted resolution
FAQ file	faq.md in the product folder — grep -i "keyword" faq.md before drafting from scratch
Log the run	append one summary line to support-log.md (date, counts, escalations)
Spot repeat questions	grep -c -i "password reset" faq.md and eyeball the batch for twins
Trend check	tail -5 support-log.md — is P1 volume rising week over week?

Procedure

1. Precheck — confirm the message source (file export or connected tool) and locate faq.md and the brand-tone notes if they exist (ls, then read_file). No tone notes? Default: warm, direct, first-person, no corporate filler — and ask the user once to confirm.
2. Triage pass — read every message before answering any. Tag each P1–P4. Within a bucket, oldest first. P1s surface to the user immediately, even before drafting the rest.
3. Escalation gate (before drafting) — flag for human handling, with a one-line summary and NO drafted resolution, anything that involves: refund or chargeback requests, legal threats or lawyer language, data/privacy incidents, or clearly angry customers (profanity, "scam", threats to post publicly). For these you may draft only a holding line: "Thanks for flagging this — {Name} is looking at it personally and will reply today." The user decides everything else.
4. Draft replies — for the rest, check faq.md first (grep the topic) and reuse the canonical answer; otherwise write fresh. Every draft: acknowledge the specific issue in the first line, give the answer or the exact next step, one question max back to the customer. Match brand tone. Never send — save all replies as drafts labeled with the customer name.
5. FAQ capture loop — any question asked twice-or-more in this batch (or once here plus once in faq.md history) gets its best-drafted answer appended to faq.md under a ## {Question} heading. This is how next week's batch gets smaller.
6. Deliver — output the template: escalations first, then drafts by severity, then FAQ updates, then the log line. Remind the user: drafts only, nothing sent.

Output Template

# Support Batch — {Date} ({N} messages)

## 🚨 Escalated to you ({n})
- {Customer} — {refund/legal/angry} — {one-line summary} — holding draft ready: yes/no

## Reply drafts ({n})
### P1 — {Customer}: {issue}
> {draft reply}
### P2/P3 ...

## FAQ updated
- Added: "{question}" (asked {n}x)

## Numbers
P1 {n} · P2 {n} · P3 {n} · P4 {n} · escalated {n} · FAQ +{n}
Nothing has been sent — review drafts and fire when ready.

Pitfalls

• Agent resolves a refund itself ("I've processed your refund") — money moved, or worse, a promise the business can't keep. Recovery: money is always human territory; strip any commitment language from drafts and route to the escalation list.
• Tone whiplash — templated corporate boilerplate to a customer who wrote casually, or breezy jokes to a furious one. Recovery: mirror the customer's register; re-read their message before finalizing each draft.
• Answering messages in arrival order — a P1 outage waits behind six P4 thank-yous. Recovery: triage the whole batch before drafting anything; the template forces severity order.
• FAQ never updated — the same "how do I reset my password" gets hand-drafted forever. Recovery: step 5 is not optional; if a question repeats and faq.md didn't grow, the run isn't done.
• Draft sent by accident via a connected tool — recovery: only ever use draft-creation calls, never send calls; if the tool lacks a draft mode, output replies as text for the user to paste.

Verification

• Every message in the batch appears exactly once (escalated or drafted) — counts add up
• All refund/legal/angry messages are in the escalation list with no resolution promised
• Zero messages were sent; all replies exist as drafts or paste-ready text
• Repeat questions appended to faq.md (diff shows the additions)
• Summary line appended to support-log.md
• Each draft's first line acknowledges the customer's specific issue (no generic openers)