Customer Support Reply

# Task
Draft a customer support reply. Use ONLY the provided policies and facts.

# Inputs
- Customer message: {customer_message}
- Policies or facts: {policies_or_facts}
- Desired resolution: {desired_resolution}
- Tone: {tone}

# Anti-hallucination rules
1. Never promise outcomes not supported by the policies.
2. Never invent features, timelines, or workarounds.
3. If you don't have enough information to resolve, say so and list what's needed.

# Reply framework
1. **Acknowledge** (1 sentence): Show you understand the issue. If the customer is frustrated, validate the emotion before addressing the substance.
2. **Respond** (2-4 sentences): Address the core issue using policy or facts. If the answer is "no":
   - State what you CAN do, not just what you can't.
   - Explain the reason briefly (policy, technical limitation).
   - Offer the closest alternative.
3. **Next step** (1 sentence): Exactly one clear action — what YOU will do or what the CUSTOMER should do, with a timeframe if applicable.
4. **Close**: Brief, warm, matches the tone.

# De-escalation rules (apply when customer is upset)
- Never use "unfortunately" more than once.
- Never blame the customer or imply they should have known.
- Lead with empathy, then move to resolution quickly.
- If escalation is warranted, say so directly: "I'm escalating this to [team] who can [action]."

# Output
1. Reply draft
2. Internal note (1-2 sentences: what to monitor, escalation flag if needed)
3. Missing information (if resolution requires data you don't have)

<context>
  <customer_message>{customer_message}</customer_message>
  <policies_or_facts>{policies_or_facts}</policies_or_facts>
  <desired_resolution>{desired_resolution}</desired_resolution>
  <tone>{tone}</tone>
</context>
<task>Draft a support reply using ONLY the provided policies and facts.</task>
<instructions>
  Never promise unsupported outcomes or invent features. If info is missing, say what's needed.

  Structure: Acknowledge (validate if frustrated) > Respond (use policy; if answer is no: state what you CAN do, explain why, offer alternative) > Next step (one action with timeframe) > Close.

  De-escalation: max 1 "unfortunately", never blame customer, lead with empathy, escalate explicitly when warranted.
</instructions>
<output>
  Return: reply draft, internal note (1-2 sentences), missing information (if any).
</output>

Customer message: {customer_message}
Policies or facts: {policies_or_facts}
Desired resolution: {desired_resolution}
Tone: {tone}

Rules: only provided policies, never promise unsupported outcomes, list missing info if needed.

Structure: Acknowledge (validate frustration) > Respond (policy-based; if no: say what you CAN do, explain, offer alternative) > Next step (one action, timeframe) > Close.

De-escalation: max 1 "unfortunately", never blame customer, escalate explicitly when needed.

Based on the entire content above, return: reply draft, internal note (1-2 sentences), missing information (if any).