ChatGPT Prompts for Customer Support Replies That Don't Promise What You Can't Deliver

When you're a solo founder, you are the support team. Every "Where's my download?", "Can I get an invoice?", and "This doesn't work" lands in your inbox, usually while you're building something else. ChatGPT can turn a two-line note into a calm, clear reply in seconds, which is exactly why it's tempting to paste a customer's message and send whatever comes back.

The risk is not the tone. It's the facts. A model that doesn't know your policies will write plausible ones: a refund window you never offered, a delivery date you can't hit, a feature that doesn't exist. This guide shows a safer workflow: write a short policy brief once, draft replies that use only those facts, and check every draft before it goes out. You'll get three copy-paste prompts and a short checklist.

Why "just reply to this" goes wrong

ChatGPT is built to give helpful, confident answers. If your prompt doesn't say what's true for your business, it fills the gap with what's typical for businesses in general. In a support reply, a typical answer stated as your answer becomes a promise.

That promise can stick. In Moffatt v. Air Canada (2024), a support chatbot on Air Canada's website told a customer he could apply for a bereavement fare after travel; the airline's own policy page said otherwise. The airline argued it couldn't be held liable for what the chatbot said. The tribunal disagreed: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot" (Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149).

That was a Canadian small-claims decision about a chatbot answering customers directly, not legal advice for your situation. But the lesson carries over to a solo founder pasting AI drafts into email: whatever you send, you said.

What to gather first (about 30 minutes, once)

Write a one-page "support brief" you can paste into every session:

• What you sell, in one line, and how it's delivered (download link, account, shipped item, call).
• Your real policies, copied word for word from your site: refunds, delivery, invoices, cancellation, data deletion. If you don't have a policy for something, write "no policy, ask me."
• Known issues and their fixes, e.g., "download link expired: I resend from the dashboard."
• What only you decide: refunds, discounts, exceptions, legal or billing disputes, anything angry or public.
• Your voice: two or three replies you've written that you liked.

Before pasting customer messages, remove what the model doesn't need: full names, email addresses, order IDs, payment details. OpenAI's own help center notes that "On personal workspaces with ChatGPT Free, Plus, or Pro, data sharing is enabled by default," meaning conversations may be used to train models unless you turn off "Improve the model for everyone" in Data controls (OpenAI Help: Data controls in ChatGPT). Check that setting, and mask personal details either way.

Prompt 1: Turn your policies into a support brief

Start a dedicated chat or project and set the rules once.

You are helping me, a solo founder, draft customer support replies. You never send anything; I review every draft.

Here is everything you may treat as fact:
---
Product (one line): [WHAT IT IS + HOW IT'S DELIVERED]
Policies (copied from my site): [REFUNDS / DELIVERY / INVOICES / CANCELLATION / DATA]
Known issues and fixes: [ISSUE -> FIX]
Things only I decide: [e.g., refunds, discounts, exceptions, legal or billing disputes]
---

Rules for every draft in this chat:
1) Use only the facts above. If a question needs a fact that isn't here, do not guess. Write [NEED: what's missing] in the draft.
2) Never offer refunds, discounts, credits, deadlines, or exceptions unless the policy above says so.
3) If a message touches anything in "Things only I decide", start your answer with ESCALATE and explain why in one line.
4) Never blame the customer or make up a reason for a problem.

First, read the policies and list:
a) Any policy that is unclear or that two customers could read differently.
b) The 5 questions customers are most likely to ask that these facts do NOT answer.

Good output: a short list of gaps in your own policies. Fix those on your site first; every gap is a reply you'd otherwise improvise.

Prompt 2: Draft the reply (calm, short, one next step)

Use this for each message, in the same chat so the brief and rules still apply. The rules can drift in long chats, so start a fresh chat with the brief every so often.

Draft a reply to this customer message using only the support brief and rules above.

Customer message (personal details removed):
---
[PASTE MESSAGE]
---
What I know about this case: [e.g., bought 2 days ago, download link sent, no previous emails / or "nothing yet"]
Channel: [EMAIL / CHAT / SOCIAL DM]
Customer mood: [neutral / confused / frustrated / angry]

Write:
1) The reply, under 120 words: acknowledge the specific problem in the first sentence, answer it or ask for the one piece of information you need, and end with one clear next step.
2) If the customer is upset: apologize once, without excuses, and do not repeat "I understand".
3) Below the draft, list every factual claim in it and the line of the brief it comes from. Mark anything without a source as UNSUPPORTED.

Avoid: "We apologize for any inconvenience", "rest assured", "valued customer", exclamation marks, emojis.

The claim list in step 3 is the important part. It takes you ten seconds to scan, and it's where invented policies show up.

Prompt 3: Pre-send check (and turn repeats into saved replies)

Run this on the draft you're about to send, especially for refund, billing, or angry messages.

Check this support reply before I send it.

Support brief: [PASTE OR "as above"]
Customer message: [PASTE]
My edited draft: [PASTE]

Check and report:
1) PROMISES: any refund, discount, credit, date, deadline, or feature mentioned. Is each one backed by the brief? Quote the line or mark NOT BACKED.
2) TONE: anything that sounds defensive, sarcastic, blaming, or overly corporate. Suggest a plainer line.
3) GAPS: did the reply answer what the customer actually asked? If not, what's missing?
4) PRIVACY: does the reply reveal anything about other customers, internal tools, or details the customer didn't give us?
5) RISK: should this go to me for a decision before sending (money, legal, public complaint)? Yes/No and why.

Then, if this is the kind of question that will come up again, write a reusable saved reply with [PLACEHOLDERS] and a one-line note on when to use it.

Over a few weeks the saved replies become your FAQ. Questions that keep coming up are also a signal to fix the product page or onboarding; the onboarding emails guide covers that side.

A quick checklist before you hit send

• Every fact in the reply is in your support brief or you checked it yourself (order status, delivery, account).
• No refund, discount, or date appears that you didn't decide.
• The first sentence names the customer's actual problem.
• One clear next step, and the customer knows what happens after it.
• No personal data in the chat that the draft didn't need.
• You'd be fine seeing the reply screenshotted in public.

Common traps

• Letting the model answer order-specific questions. ChatGPT can't see your Stripe dashboard or inbox. It can phrase "your download link is below"; you supply the link.
• One giant prompt per ticket. Set the brief once, then reuse it. Shorter prompts, fewer mistakes.
• Over-apologizing. Three apologies read as a script. One, then the fix.
• Auto-sending. For a solo business, reviewed drafts are the right default. Automate only after you've reviewed enough drafts to know where they go wrong, and keep money and disputes manual.

The short version

1. Write a one-page support brief with your real policies, fixes, and the decisions only you make.
2. Use Prompt 1 to find the gaps in those policies and set the rules.
3. Use Prompt 2 for each message: short reply, one next step, and a list of every claim with its source.
4. Use Prompt 3 to check promises, tone, privacy, and risk, and save repeat answers.
5. Read every draft before it goes out. You're responsible for it either way.

ChatGPT makes support replies faster and calmer. Knowing what you've actually promised is still your job. If customers keep asking the same pre-sale questions, the landing page copy guide will help you answer them before they email.