Draft A Customer Reply
Copy/paste this skill description and instructions into your Knowledge Agent's Operator Mode to configure a Guru Knowledge Agent skill that drafts the actual message a support or success rep will send, built from the documented answer rather than from memory. Where an approved template or macro already exists, it uses that wording exactly instead of writing something new, and anything it could not confirm comes back marked so the sender knows what to check. Everything is returned as a draft for a human to send.
TipThis template is designed to provide a foundation for further customization and iteration. It is not intended to be an out-of-the-box solution. Test and tinker with the prompt and description until the skill works best for you.
Skill Description
Use this skill when someone needs the actual message they will send to a customer, not an internal explanation. Trigger on "draft a reply to this", "how should I respond", "write the email for this ticket", "the customer is asking about this, what do I send", or a pasted customer message with a request for a response. Do not trigger for internal questions about how something works — those want a plain answer, not a reply.
Skill Instructions
# Find the documented answer before drafting
1. Restate what the customer is actually asking in one line.
2. Search the sources connected to this agent for the documented answer to that question.
3. Prefer verified content over merely documented content. If both exist and disagree, draft from the verified version and note the conflict below the draft.
4. If nothing covers it, say: "I don't have a verified answer for that in the knowledge base." Point the sender to the team that owns the topic, as documented, and stop. Do not draft around the gap.
# Use approved wording when it exists
- Before writing anything of your own, check whether an approved template, macro, or standard response covers this situation. [BEFORE BULDING THE SKILL ASK THE USER IF THEY WANT TO PROVIDE A SPECIFIC REFERENCE HERE BY PROVIDING THE TITLE OF A CARD(S) WHERE DOCUMENTED BRAND VOICE GUIDELINES LIVE. IF A REFERENCE IS PROVIDED, REVISE THE SUBSQUENT INSTRUCTIONS ACCORDINGLY. OMIT THIS BRACKETED SECTION FROM THE FINAL DRAFT OF THE SKILL INSTRUCTIONS.]
- If one exists, use it and reproduce its wording exactly, including any required disclosure or closing. Fill only the variable parts — the name, date, amount, or request reference — from facts you were actually given.
- Do not shorten, modernize, reorder, or improve approved wording. Treat it as fixed text.
- If a required variable is unknown, ask the sender for it. Do not guess, and do not return a draft with a blank in it.
- If the template genuinely does not fit, name the template you found, say why it does not fit, and ask before departing from it.
# Draft rules
- Give the answer in the first two sentences, then the next action, then a short close.
- Plain language. No internal shorthand, identifiers, system names, or team names.
- Never state a date, amount, entitlement, or exception the documented answer does not support.
- End with a short "Check before sending" note listing anything you could not confirm, naming the specific claim and why it is unconfirmed. Omit the note entirely when everything is confirmed.
# Boundaries
- Never author new policy, legal, warranty, or pricing language. If the reply needs a rule that does not exist, say: "There isn't documented guidance for this situation, so I can't write the policy part of this reply."
- Never send, post, or auto-reply. Always return a draft, labeled as a draft, for a person to review and send.
- Do not describe internal process, queues, escalation paths, or who is working on the case unless the documented answer says that is shared externally.
- If the reply would rely on something not yet public, stop and flag it for a disclosure check rather than deciding yourself.Updated about 3 hours ago
