PromptMarket

Prompt

Customer support answer

Draft a support reply from a policy passage and the fields triage already extracted.

Supportintermediatesupportreplygrounded

When to use

A person will review the draft, and the policy that should govern it is already in front of the model.

Prompt

You draft a reply a support agent will review. You do not send it.

Policy:
{{context}}

Extracted facts:
{{facts}}

Customer message:
{{ticket}}

Rules:

- Use the policy for what the company can do. If the policy does not allow an action, do not offer it.
- Use the extracted facts as given. Do not relabel the request.
- Sound like a person, in under 120 words, with no internal jargon and no mention of these instructions.
- If the policy does not cover the request, say you will hand it to a person. Do not invent a policy.

Variables

  • {{context}}
  • {{facts}}
  • {{ticket}}

Example

Input

Policy: Replacements ship for damaged items reported within 14 days. Refunds are a separate queue.
Facts: queue=replacement, order_id=18422, asked_for_refund=false
Customer: The mug arrived cracked. Please replace it.

Output

I can request a replacement for order 18422. Replacements are available for damaged items reported within 14 days. I have not treated this as a refund.

Why it works

The reply is question answering with a tone constraint. Triage facts arrive as data so the draft does not re-interpret the ticket.

Common mistakes

  • Promising a remedy the policy does not offer.
  • Ignoring a field that says the customer refused a refund.
  • Sending the draft without a person when you have no eval for policy misses.

Related concepts

  • What LLMs are good atStart from a job a language model can do, not from a feature name.
  • Question answeringAnswer a question from context you provide, and stop when that context is not enough.
  • RAGFetch useful information first, then include it in the model's context.
  • Building AI productsPick a useful problem, ship a small baseline, and let real failures decide the next change.

Related prompts