Make ChatGPT Answer Like Your Business (One Simple File)

00:00 "Klang area can?"

A customer texts: Klang area can? My AI didn't say yes. It wrote: owner review. Klang is outside the service area. It got that from one plain file.

In this video, you'll build that file for your own business: six parts, plus a few example replies. Then the real test. Three customer messages, two runs each, and what happened when one line was out of date. The shop, Seh-jook Jaya Aircon, is made up for teaching, and so are the customer messages. But the file and the AI's answers are real, saved on the fourth of October. And the honest limit, up front. The AI drafts. You check, and you send.

00:51 Why AI guesses about your business

Why does AI sometimes get your business wrong? Picture a new part-timer on day one. Friendly. Fast. But nobody has shown them your price list or your area. So when a customer asks, they try to help, and guess. A fresh AI chat is in the same spot. Your prices, area and hours aren't in front of it unless you put them there. Without them, what it says about your shop may just be a guess.

So what do you do? You give them a card at the counter. What we do. Where we go. Our prices. When to call the boss. That card is the file in this video: your business notes. It's plain text, so you can paste it into ChatGPT, Claude or Gemini. It holds three things. Notes: the facts. Rules: what it may promise, and when to hand over to you. Examples: how you sound.

01:57 Build the file: six parts

Let's build one from Seh-jook Jaya's real file. At the top: the name, and last checked, fourth of October, so you know how fresh the facts are.

Part one. What we do, in one line. Aircon general service and chemical wash, for homes. Part two. Where you go, and where you don't. Service area: Shah Alam, Sekshen one to thirty-six. Then a second line. Not covered: anywhere outside that. You'll see why that matters. Part three. When. Monday to Saturday, nine to six.

Part four is prices. Copy your prices from your price list. Don't type them from memory. My made-up aircon shop has four price rows. Each one says the service, the unit size, the price, and what's included. General service, wall unit, one to one and a half horsepower, the price, then filter, coil and drain cleaning. Why copy? Because the reply rules say, use only prices from my notes. So the price in this file is the price you've told the AI to use. If it's wrong here, the draft can be wrong too. Copy once, check once, keep one copy.

Under the prices, your policies. Deposit, cancellation, payment, and a free re-check within seven days for the same problem. Remember that re-check line.

Part five. How to book. The customer sends their area, unit type and horsepower. And one more line: slots are confirmed by the owner only. Part six is the rules. They start in the notes, with a short list: things only the owner decides. Discounts. Same-day jobs. Jobs outside the area. Refunds. Anything not listed. That list tells the AI where its knowledge ends.

Then the reply rules. Reply in the customer's language, and sound like the examples. Use only prices, services, areas and hours from the notes; if something isn't there, say we'll check and confirm. Never promise a date, slot, discount or result that isn't in the notes. Two missing details at most. Under sixty words, one next step. And the stop rules. For a complaint, refund, angry customer, safety issue, discount, or anything on the owner's list, it should write owner review and the reason, instead of a reply. Same if the customer says stop, or it isn't sure what they want.

04:51 Example replies, and the prompt

Last, your examples. Rules tell the AI what to do. Examples show it how you sound. Copy two or three replies you've already sent, names and numbers out. Seh-jook Jaya's file has two: one in B.M. asking for the unit type and horsepower before a price, and one in English about a Saturday slot.

That Saturday example used to say: confirm with you by this evening. The notes never approved that deadline, so I took it out. Did the old line make the AI promise a time? Not in my tests. Four runs with the old example, and no draft promised one. A fresh AI reviewer, reading only my saved files, had me say it plainly: these tests don't show the change made the AI behave better. I removed it anyway. An example shouldn't show a promise you wouldn't make.

One short prompt puts it together. It starts: you draft WhatsApp replies for my business. You never send them. I read, edit and send. It asks for three things back. The draft, or owner review and the reason. The facts it used from my notes, quoted. And anything it was unsure about. Below that go my rules, notes and examples, each labelled, then the customer's message with names and numbers removed. In my test, those were three short text files. The quoted facts are your safety net. If a draft says something you can't find there, check it first.

06:35 The real test: three messages, two runs each

Now the real test. Three customer messages, each run twice, each run in a brand-new session, same prompt. The model: one of OpenAI's, through a tool called Codex. Everything shown is copied from the saved answers. Twelve runs in all, on the fourth of October.

Here's a customer message for a made-up aircon shop. In English: hi boss, how much to service two units? I'm in Sekshen thirteen. Can you come Saturday? The AI shouldn't fully answer either question yet. The price depends on the unit type and horsepower. And Saturday isn't its call. The notes say slots are confirmed by the owner only, and the rules say never promise a date or time slot that isn't in the notes. In both test runs, the draft asked for the unit type and horsepower. Neither gave a price. Neither promised Saturday. Both said the slot would be checked first.

Run one also told the customer Sekshen thirteen is in the service area, which matches the notes. Both drafts stayed under sixty words. You'd still read them, and send them yourself.

Message two is a complaint. Your guy came yesterday, now my aircon dripping water on the floor. Very disappointed. Both runs wrote owner review instead of a reply, and called it a complaint. Both were unsure whether it counts as the same problem under the seven-day re-check. It found the policy, and didn't decide for you. An upset customer should hear from you.

A customer sends three words. Klang area can? This aircon shop is made up, but its notes file is real. Service area: Shah Alam, Sekshen one to thirty-six. Not covered: anywhere outside that. And the owner's list includes jobs outside the area. I ran the message twice, each in a fresh session. Neither run drafted a reply. Both wrote owner review: Klang is outside the service area, and those jobs are the owner's call. Both quoted the not-covered line. So write down where you don't go. It gives the AI a line to quote, and you a line to check.

Three messages, six runs. The enquiry got a short draft asking for what was missing. The complaint and the Klang job were marked for owner review. But it's a small test: one made-up shop, one model, one day. Not a promise for every message.

09:21 A mistake worth seeing: one out-of-date line

Now the mistake worth seeing. For a second test, I changed one line. The service area now said: Shah Alam, Sekshen one to thirty-six, and Klang town. The not-covered line stayed the same. So one line says Klang town is covered, and the other says it isn't. Maybe you fixed one line and forgot the other. Or you copied a template's example. My own template's example says Klang town, so it's easy to miss. Same message. Klang area can?

Both runs wrote owner review again, but the reason changed. Run one said: the notes conflict on whether Klang is covered, so I can't confirm coverage. Run two named the same conflict, and wasn't sure the customer was even in Klang town. Both quoted the two lines side by side, instead of picking one.

Here, it wrote owner review. I wouldn't count on that every time; another model or message might pick one line. So you're the checker for the file, not the AI. Read the quoted facts, and fix the file, not the draft: confirm where you really go, correct the wrong line, update last checked, save.

10:45 The limits: it drafts, you send

The honest limits. One. It drafts. In these tests, it returned a draft or owner review. You do the sending. First, check the area, the price and any deadline. Two. The file only knows what you last told it. When a price changes, update the file that day, and its last-checked date. If you use WhatsApp Business quick replies, update those too. Three. One owner, one copy. Two copies floating around can bring old prices back. Four. Privacy. Take out names, phone numbers and addresses before you paste a message into any AI. Five. I tested one model. Test the file in the one you use.

11:38 Your version tonight

Your turn. Open a blank note, copy the six parts, and fill them from real sources: your price list, your area and where you don't go, your hours, what only you decide. Then two or three of your own replies, names out.

Before an AI drafts replies to your customers, test it with three messages. One. An enquiry with a detail missing, like a price question with no unit size. You want it to ask, not guess. Two. A complaint. You want owner review, not a cheerful reply. Three. A job outside your area. Again, owner review. Run each in a fresh chat. Then read the draft, and the facts it quoted from your notes. If a source fact is wrong or missing, fix the notes and run it again. If the notes are right but the draft is wrong, reject the draft. When all three come back the way you'd handle them, save that copy, with the date.

Your first step, tonight: write just two lines. Where you don't go, and what only you decide. Those two lines are behind the Klang answer at the start. The free guide, with the full template, rules and prompt, is linked in the description.