customer-reply

SKILL.md

Customer Reply

Drafts a brand-consistent, factually grounded reply to an inbound customer message. Three jobs: (1) figure out what the customer is actually asking, (2) ground the answer in the team's FAQ to avoid making things up, (3) write it in the team's voice — not Claude's default voice.

When to trigger

• The user pastes a customer email, ticket, support chat, or DM and asks for help replying.
• The user says any of: "draft a reply to this", "respond to this customer", "what should I say to this person", "answer this ticket", "help me write back to this user".
• The user shares a screenshot or text from Intercom, Zendesk, Front, HelpScout, Crisp, Plain, or similar support tools.

When NOT to trigger

• The user asks for marketing copy, sales emails, or outbound outreach (use a different skill).
• The customer message is in a language that has no entry in brand_voice.md for that language — surface that gap to the user instead of guessing.
• The message contains threats, legal requests, refund disputes over €1,000, or urgent safety issues — escalate to the user, do not auto-draft.

Workflow

1. Read the inputs. Open brand_voice.md and faq_snippets.md from this skill's folder. If either is empty or templated, tell the user what's missing before drafting.
2. Classify the message. Pick exactly one category: bug, how-to, billing, feature-request, feedback, escalation, other. Print the category before the draft.
3. Find grounding. Search faq_snippets.md for relevant entries. Quote the FAQ ID(s) you used (e.g. [FAQ-12]) inline next to claims. If no FAQ entry covers the question, label the answer "ungrounded — verify before sending" and don't fabricate.
4. Draft the reply. Match the voice rules in brand_voice.md. Default length: 2-4 short paragraphs. No emoji unless brand_voice.md permits them. Sign with the placeholder {{agent_name}} — the user fills it in.
5. Add a "next-step" line. Always end with a clear next step (e.g. "Just hit reply if that didn't fix it" or "I've opened ticket #1234 to track this"). Customer-support reply quality lives or dies on the next-step line.
6. Output the reply, then a meta-block. Format:

   <reply>
   ...the draft...
   </reply>
   
   <meta>
   - Category: bug
   - FAQs used: FAQ-12, FAQ-04
   - Confidence: high | medium | low
   - Flags for human: [list anything you weren't sure about]
   </meta>

Hard rules

• Never invent facts about the product. If the FAQ doesn't cover it, say so in <meta> flags.
• Never promise a refund, timeline, or feature — those are the user's calls. Defer with: "Let me check with the team and get back to you within 24h."
• Match the customer's language. If they wrote in Dutch, reply in Dutch. If brand_voice.md has no Dutch tone profile, fall back to "warm, professional, no jargon" and flag it.
• Don't apologize gratuitously. One sincere sorry where warranted, not three. Apology spam reads as performative.
• Skip "Hope this helps!" It signals the writer didn't read the question. Use a real next-step line instead.

brand_voice.md

Edit this file once. The skill reads it every time it drafts.

Default tone profile

• Warmth: medium-warm. We're friendly but not chatty.
• Formality: casual-professional. First names. No "Dear Sir/Madam".
• Hedging: low. Say what you mean. "Yes, that's a bug." not "It seems like there might possibly be an issue."
• Apology budget: at most one apology per reply, and only when we genuinely caused friction.
• Emojis: none.
• Sign-off: "— {{agent_name}}"

Phrases we use

• "Got it —"
• "Quick fix:"
• "Let me dig into that and get back to you within {{timeframe}}."
• "Two options here:"

Phrases we never use

• "Hope this helps!"
• "Reaching out to follow up..."
• "I completely understand your frustration."
• "Per my last email"
• "As mentioned"
• "Synergy", "leverage" (verb), "circle back"

Voice tests

A reply passes if a colleague reads it and says: "yeah, that sounds like us." A reply fails if it sounds like a chatbot, a lawyer, or a yoga retreat.

faq_snippets.md

The skill grounds answers in this file. Add or replace entries with your own. Keep IDs stable so old replies stay traceable.

FAQ-01 — Password reset email not arriving

• Most common cause: email landed in Promotions/Spam. Search "{{product_name}}".
• Second most common: typo in the address on signup. Customer can re-register with the correct address; old account auto-merges within 10 minutes.
• If neither: reset emails go through Postmark. Latency >2 minutes is rare. Have them try a Gmail/Outlook alt to rule out their server filtering us.

FAQ-02 — How do I cancel my subscription?

• Settings → Billing → "Cancel subscription". Effective at end of current period.
• We don't offer pro-rated mid-cycle refunds, but we do offer a free month if there's a specific issue we can fix.

FAQ-03 — Can I get an invoice with my company VAT number?

• Yes. Settings → Billing → "Add company details" before next billing. Past invoices can be regenerated by emailing finance@{{company_domain}} with the invoice number.

(4 more entries in the ZIP: 2FA codes, data export, pricing changes, API rate limits.)

examples/01_input.md → examples/02_output.md

Input: a customer ticket from "Sara K." reading "I've been trying to log in for the last 30 minutes and the password reset email never arrives. I've checked spam. This is incredibly frustrating, I have a meeting in 20 minutes."

Output the skill produces:

<reply>
Hi Sara,

Got it — sorry the reset email is dragging. Two quick things to try:

1. Search your inbox (not just Spam) for the product name. Gmail and
   Outlook sometimes route reset mails into Promotions or a category tab.
2. If that turns up nothing, try requesting a reset to a Gmail or
   Outlook address you control. That rules out a filter on your work
   mail server, which is what we usually see when the email truly
   never arrives.

If neither works, reply with the email you signed up with and I'll
re-send it from our end and confirm delivery.

— {{agent_name}}
</reply>

<meta>
- Category: how-to
- FAQs used: FAQ-01
- Confidence: high
- Flags for human: customer mentions a meeting in 20 minutes — if reply
  isn't sent in <5 min, this should be escalated to a human or a phone
  call. Consider auto-prioritizing tickets with explicit time pressure.
</meta>