Support replies with AI is a Skill: one saved prompt that reads your help pages and drafts a tone-matched answer to each customer question, and setting it up takes one sitting of about an hour. After that, each reply takes a minute or two, because you paste the question, read the draft and send it yourself. The AI drafts from your help center, in your voice. You decide what is true and you press send. That split is the whole job, and this page shows how to set it up today with the AI you already pay for.
- Shape
- Skill
- Sitting
- one sitting
The shape: Skill
This job is a Skill because it is one prompt and the material under it. There is nothing to connect and no chain of steps. The quality of every draft depends on two things: what you load into the prompt, and the rules you give it about what it may not say.
The AI is good at this part of support. It reads a long policy page faster than you do, it finds the one line that answers the question, and it can match the tone of replies you already wrote. It also stays calm with a customer who is not.
What it is bad at is knowing anything you did not give it. It does not know that you changed your cancellation window last month, that this customer has been with you for six years, or that the part they want is on backorder. A Skill set up well makes that limit visible: the draft says when the answer is missing instead of guessing.
The payoff is consistency. Without a saved prompt, each reply starts from whatever mood you are in at 9 at night. With one, every customer gets the same accurate answer in the same voice, and your edit is a quick check instead of a rewrite.
What you give the AI
Gather these into one document before you open the chat. Label each piece so the AI knows what it is reading.
- Your help pages or FAQ: copy the text of every page a customer might need, such as shipping, returns, booking, cancellations, pricing questions and how your service works.
- Your written policies: refund rules, warranty terms, the cancellation window, anything you would point to in a dispute.
- Three to five support replies you already sent: real ones from your sent folder, including one to an upset customer and one that said no.
- Your brand voice guide, if you have one: if not, build one first with the brand voice guide job, or let the sent replies stand in for it.
- A list of things the AI must never do: promise a refund, offer a discount, admit fault, give medical, legal, tax or financial advice, or quote a date you have not confirmed.
- How customers reach you and how fast you answer: the inbox, the phone number, your hours, and the response time you can actually keep on your worst week.
- The sign-off you use: your name, the business name, or both.
Leave out anything that is no longer true. An old shipping page in the material will come back as a confident wrong answer. If your help pages are thin, that is fine for now; the prompt will tell you which questions they fail to cover.
What comes back
Save this prompt in a project or in the instructions of a custom assistant, with your material pasted below it, so it loads every time. Then paste each customer message as a new request.
You draft customer support replies for a small business.
A person reads every draft and sends it. You never send anything.
Answer only from the HELP PAGES and POLICIES below.
If the answer is not in them, do not guess. Say so in a note above the draft,
and write a short holding reply that says we are checking and when the customer
will hear back, using the response time given below.
Match the tone of the SENT REPLIES and VOICE GUIDE below: their length, greeting and
how they handle bad news. End with the SIGN-OFF below. Write as "we" unless the samples show otherwise.
Never promise a refund, a discount, a credit, an exception or a date that the
policies do not state. Never admit fault. Never give medical, legal, tax or
financial advice. If the message asks for any of these, flag it for the owner.
For each customer message, return:
1. A one-line read of what the customer wants and how they feel.
2. The reply, answering the question in the first sentence, giving one next
step, and stopping. Short enough to read on a phone without scrolling.
3. Sources: the help page or policy each fact came from.
4. Flags: anything the owner must check or decide before sending.
HELP PAGES:
[paste]
POLICIES:
[paste]
SENT REPLIES:
[paste]
VOICE GUIDE (skip if you have none):
[paste]
SIGN-OFF:
[paste]
NEVER LIST:
[paste]
CONTACT, HOURS AND RESPONSE TIME:
[paste]What comes back for each question is four short parts: the read, the reply, the sources and the flags. The reply is what you send. The other three are for you, and they are what makes each draft fast to check. The sources tell you in a glance whether a fact came from your pages or from the AI's idea of a typical business. The flags tell you where your judgment is needed before the reply goes anywhere.
Over a week, the flags also become a list. Every "not in the help pages" note is a question your customers ask that your pages do not answer. Write the answer once, add it to your help pages and to the prompt material, and that question drafts itself from then on.
How to judge it
Read the draft once and reject it if you see any of these:
- A fact with no source: a delivery time, a fee or an exception appears in the reply but not in the sources list, so the AI filled a gap.
- It answers a different question: the customer asked when the order ships and the reply explains how returns work.
- The answer is buried: two sentences of sympathy come before anything useful.
- It promises what you did not authorize: "we'll refund that right away" or "we'll make an exception this time".
- It sounds like a call center: "we apologize for any inconvenience", "your satisfaction is our priority", or a reply three times longer than anything you have sent.
- It misreads the mood: a cheerful tone to someone who is clearly upset, or a long apology to someone who asked a simple question.
- It gets the customer's details wrong: the wrong name, the wrong product, the wrong date pulled from their message.
Then run the real test before you trust it. Take ten customer questions you already answered by hand this month. Paste each one in and compare the draft with what you sent. If you would send most drafts with small edits, the setup works. If you would rewrite them, look at what you changed: usually a missing policy, a tone rule, or a word you never use, and add it to the material.
What stays with you
The AI drafts. You decide. That line does not move, and here is what sits on your side of it.
The facts. The AI only knows what you pasted. Whether the order shipped, whether the appointment slot is still open, whether this customer already got a credit last time: those live in your systems and your head, and you check them before you send.
The exceptions. A refund outside the window, a free redo, a waived fee: these are business decisions about a real relationship. The AI can draft the words once you decide, but the decision is yours.
The hard messages. A complaint that mentions a lawyer, an injury, a safety issue, a regulator or a staff member is not a support reply. Handle it yourself, and get advice where it is needed. If the complaint is a public review, the guide to negative Google reviews covers what to check before anything goes public.
The send button. Every reply goes out from your inbox, under your name, after you read it. That minute of reading is cheap insurance against a wrong answer that a customer screenshots.
The upkeep. When a policy changes, the material in the prompt has to change the same day. An AI working from last season's pages will answer with last season's rules, politely and with confidence.
An editorial example
This is an invented business, used only to show the job. Say you run a small bike repair shop with online booking. You paste in your booking page, your pricing page, your warranty policy, four replies from your sent folder and a never list that includes "quote a repair price without seeing the bike".
A customer writes that their brakes squeal and asks what it will cost and whether you can see them Saturday. The draft comes back. The read says they want a price and a time, and they sound mildly worried. The reply says squealing brakes are a common fix, that you price repairs after a free look at the bike, and points them to the booking page for Saturday slots. The sources list your pricing page and your booking page. The flag says Saturday availability is not in the material, so you check the calendar before you send.
You check, Saturday is full, and you change one line to offer Monday. That is where the example stops. Nothing in this example is a real business or a real outcome.
Where the replies go next
The questions in your support inbox are the questions your next customers have before they buy. Answering them well in support is one use. Answering them before anyone has to write in is another: the doubts that stop a sale belong on an objection handling sheet, and the everyday questions make good posts, as covered in turning FAQs into social media posts. The rest of the jobs are on the full list.
Of every job on this list, social media is the one that comes back every week, and Boomp is software that does that one for you. See ten posts made from your website, free
