# Customer support reply pack

Turn your actual support policies into a reply-drafting system. An AI assistant drafts every reply grounded in your written rules, and a human on your team approves before anything is sent. Works with any capable AI assistant — ChatGPT, Claude, or Gemini.

## What you need

- Your written policies: refunds, returns, shipping, warranty, cancellations — whatever actually decides how you respond. If they only live in someone's head, write them down first. Bullet points are fine.
- 10 past customer messages along with the replies your best person actually sent.
- An AI assistant your company allows for this data. Strip customer names and account details before pasting, or use a plan whose data terms cover support content.

## Setup

### Step 1: Extract the rules that decide replies

Go through your last 20 tickets. For each one, write down the rule that decided the answer ("orders over 30 days: store credit only"). You will end up with a short list of decision rules. That list is the engine. Expect to discover rules that were never written down anywhere.

### Step 2: Fill in the policy-to-prompt template

Copy this and replace every bracketed section with your reality:

```
You are a support reply drafter for [COMPANY], which sells [WHAT YOU SELL]
to [WHO BUYS IT].

POLICIES — these are the only rules you may use:
[PASTE YOUR DECISION RULES, ONE PER LINE. Example format:
- Refunds within 30 days of delivery: full refund to original payment method.
- 31-60 days: store credit only. Past 60 days: no refund; offer replacement at cost.
- Shipping delay: apologize, share tracking, offer nothing unless the delay
  exceeds 10 business days.]

TONE:
[2-3 RULES. Example: Warm but direct. Use the customer's first name. Never
write corporate filler like "we apologize for any inconvenience."]

ESCALATE — do not draft a reply. Respond only with "ESCALATE:" and the
reason, when the message involves:
[YOUR TRIGGERS. Example: legal threats, chargeback mentions, press or media,
anything the policies above do not cover.]

For every customer message I paste, respond in exactly this format:
1. DRAFT: the reply, ready to send.
2. POLICY USED: quote the rule above that the draft relies on.
3. FLAG: "confident" if the policies clearly cover this, or "check this"
   plus what a human must verify before sending.

Never invent a policy, discount, or promise that is not written above.
If the policies do not cover the situation, escalate.
```

### Step 3: Save it where drafting happens

Put the finished prompt somewhere reusable: a Claude Project, a ChatGPT custom GPT or project, a Gemini Gem, or a shared doc your team pastes at the start of each session.

### Step 4: Run the drafting loop

Paste a customer message. Read the DRAFT, check POLICY USED against reality, and honor the FLAG. Edit if needed, then send from your own helpdesk. The human always sends.

## QA before you trust it

Take the 10 answered tickets you collected earlier. Run each through the prompt and compare the draft against the reply that was actually sent.

- Invented policy or promise: fix the POLICIES section. It is usually a missing rule, not a broken model.
- Wrong tone: add a tone rule with a good example and a bad example.
- Missed escalation: add the trigger.

Re-run the same 10 until the drafts match what your best person would send. Repeat this check monthly. Policies drift, and the prompt has to drift with them.

## When this outgrows DIY

This setup works while a human pastes messages one at a time. It stops working when volume climbs, when replies need live order or account data, or when you want drafts appearing inside your helpdesk automatically with logging and QA on every one. That is a built system with your data connected, which is the kind of thing we install. If you get there: agentclawhq.com/book.