The Furnace Email ChatGPT Gets Wrong
Revised October 22, 2026
Can ChatGPT talk like a human?
It can write like one specific person if you show it that person. Put a few things you actually wrote into a ChatGPT project, then keep the rules in custom instructions, which hold 1,500 characters on a free account. Read every draft out loud, and cut whatever you would never say to a customer’s face.
Keep reading ↓“We appreciate your inquiry and look forward to serving your needs.” That is what comes back when you ask ChatGPT, at nine at night, to answer the woman whose furnace quit. She is reading it on a couch in a south city bungalow a few blocks from Bevo Mill, her coat still on. The sentence is grammatical. It is warm the way a hotel lobby is warm. And she can tell in the first four words that nobody who has ever knelt in front of a pilot light wrote it.
Who taught the machine to talk like that? Nobody. It is the factory default, the same polite stranger for every shop in the country, and it will sign your name to that voice until somebody teaches it better. Most owners across St. Louis try it once, wince, and go back to typing every reply at the kitchen table. Is that really the only choice?
It isn’t, and the tutorials are why it feels that way. They send you to the wrong box. They say paste your writing samples into custom instructions, but OpenAI’s help center says that box holds 1,500 characters on a Free or Go account and 5,000 on Plus, Pro, Enterprise, Business and Education. That is a couple of short paragraphs, not a voice. So where do the samples go? Into a Project, where a Free account can keep five files and where, per OpenAI, the project’s instructions override your global ones. What goes in the small box is rules. Here is a set a heating shop could paste today:
Who I serve: homeowners and landlords in the city of St. Louis and nearby.
What I do: residential heating and cooling repair.
How I sound: plain, direct, a little warm. Short sentences. First names. Never corporate.
Never invent a fact about my business: no years in business, awards, guarantees, prices or availability. If a fact is missing, leave a blank and ask me. Never write “we appreciate your inquiry” or “valued customer.”
Licenses: never write “licensed” or a license number unless I give it. Furnace permits in the city of St. Louis go through my City mechanical license, so leave a blank for it and ask me. Ask whether the job is in the city or the county before promising a permit.
Look at those last two rules, because here they carry the weight. In Missouri, the state’s Merchandising Practices Act makes any deception, false pretense, false promise or misrepresentation in connection with the sale or advertisement of any merchandise an unlawful practice, and the statute’s definitions count services as merchandise. In the city, the Building Division says mechanical permits are issued to licensed mechanical contractors certified by the City of Saint Louis, and that the permits are required to install things like boilers and furnaces. So a chatbot that signs “fully licensed and insured, family-owned for three generations” on a shop you opened last spring is not being charming. It is making claims the state and the city both know how to read, and drafting a problem under your name. This is not legal advice, but one sentence of that statute is worth reading for yourself.
Can ChatGPT talk like a human?
Not by default. Left alone it writes like the average of everyone, which is why it reads like no one. Handed a few things you actually wrote, it picks up the rhythm quickly: how long your sentences run, whether you say “heads up” or “please be advised,” whether you ever use an exclamation point. What it cannot do is hear you. It has the page, not the pilot light. So who does the last step? You do, out loud, cutting whatever you would never say to a customer’s face.
Here is the difference on that furnace email. The default says, “We appreciate your inquiry.” A project trained on your own replies might say: “Dana, that’s a rotten way to find out the igniter is shot. I can be there first thing tomorrow with the part on the truck, so it’s one trip. My City of St. Louis mechanical license number is [ask me], if you’d like to check it. Call me if it gets worse overnight.” Same facts, plus one honest blank: the bracket is the rule working, because the one fact the bot cannot know was left for you to fill in. One of those replies sounds like somebody who is coming.
Where do the samples go, and where do the rules go?
Open your sent folder. Scroll past the quotes and the “running late” texts until you hit the replies that sound like you at the end of a long day, radio on, talking to a customer you actually like. Those go in a Project. In ChatGPT you select New project in the sidebar, name it, and add material by uploading files or pasting text. OpenAI’s Projects page says Free accounts can keep up to five files in a project, Go and Plus up to 25, and Pro, Business and Enterprise up to 40. There is a catch on the Free plan: OpenAI’s upload page also says Free users are limited to three file uploads a day. Hit that wall, and the Projects page lists pasting text as an allowed way to add material.
Which replies would you want a stranger to judge you by? The one to the regular who texts at seven in the morning. The one where you said no without losing the customer. The one where you apologized and meant it. Anything a marketing person polished for you can stay out, because the model will faithfully learn the polish.
Rules go in the Project’s own instructions: select the more options menu, then Project settings. OpenAI says project instructions apply only within that project and override your global custom instructions. That gives you one voice for customer emails and a different one for social posts, each in its own Project, with your account-wide custom instructions as the floor under both.
How do the account-wide instructions get in there?
Thirty seconds at the shop counter with your phone in your hand: under Settings, then Personalization, with Enable customization switched on; on the iPhone and Android apps the menu reads Customize ChatGPT. OpenAI says they apply immediately to all chats, including existing ones. So what belongs in a box that small? Rules that never change. Notice what the heating-shop version leaves out: adjectives like “friendly but professional,” because those describe a thousand shops. Notice what it includes: a refusal list. The robot tone hides in what the model reaches for by habit, so the habits you want gone are the ones to name.
Doesn’t ChatGPT just remember how I write?
It feels like it should, after all the times you have told it about the shop. Sometimes it does, and you cannot count on it. OpenAI’s memory page says Memory does not retain every detail from every conversation, that ChatGPT decides which available information is relevant to a response, and that Memory can change as your context changes. That is fine for remembering you have a shop in the city. It is a shaky place to keep a voice you are trusting with customers. Samples in a Project stay where you put them. A memory is something the model keeps or drops on its own schedule.
The same page warns that sensitive information can appear in Memory if you share it with ChatGPT, and lists custom instructions among the sources Memory can draw on. One more reason to keep the rules short and the customer details out.
Before you paste one customer email into ChatGPT, read this first.
What stays out of the samples?
A customer’s address on Arsenal Street is not yours to give away. Customer names, phone numbers, addresses and anything with an account number all come out. Take them out before you paste, the way you would black out a check before photocopying it. The reason is on OpenAI’s data page: for individual plans, Free, Plus and Pro, it says it may use your content to train its models unless you turn off Improve the model for everyone under Settings, then Data controls. The same page says it does not use inputs or outputs from ChatGPT Business, Enterprise or Edu to improve its models by default. Even your custom instructions are covered: OpenAI says information from your use of custom instructions will also be used to improve model performance.
So the instructions box is no place for a customer’s street in Tower Grove, and the samples are no place for a last name. If you want to ask something sensitive, OpenAI says a temporary chat is not used to improve its models while it stays temporary, and it does not create memories. Passwords and card numbers never belong in any chat, on any plan.
How do you read a draft out loud and fix it for good?
Standing up, in the driveway in Affton or in the shop, at your normal speed. The places you stumble are the places the draft is lying. A tidy list of three where you would have named one thing. A phrase like “rest assured.” An exclamation point you have never once used. Cut it, then tell ChatGPT what you cut and why: “I took this out because it sounded like a brochure. Keep it this direct.” Do that a handful of times and the same tells stop coming back, because edits carry more information than adjectives ever do.
If you think faster than you type, talk. OpenAI’s Projects page lists voice among the tools you can use inside a project. Say the reply out loud the way you would in the driveway, then ask ChatGPT to fix the grammar without changing a single word or the order. You keep your real cadence, minus the “um.”
Then comes the unglamorous part. Is every number, date, hour and promise in the draft actually true? The model writes fluently whether or not it is right, and fluent is the dangerous kind of wrong.
Your first evening, in order
- Choose one kind of writing. Customer replies, review responses or the weekly post. Not all three at once.
- Collect the samples. Real messages you sent, with names, numbers and addresses taken out. A Free project holds five files, or paste them as text.
- Make the Project. New project in the sidebar, then the samples go in as sources.
- Write the rules in Project settings. Who you serve, how you sound, and the never-invent line, in a handful of sentences.
- Turn off training if you are on an individual plan. Settings, Data controls, Improve the model for everyone.
- Ask for the next real reply. Read it out loud, cut what you would never say, and tell ChatGPT what you cut.
Tomorrow morning the first reply will take you a few minutes instead of a half hour of staring, and it will have your pilot-light patience in it. The second will take less. That is the part nobody mentions: the setup is tedious exactly once, and the edits you make after that are the training. Want a person next to you for the first run? The Missouri SBDC’s St. Louis office at 120 S. Central Ave. serves the city and five counties, and 314-241-1697 reaches the office, where you can ask which AI sessions are on its calendar.
Why does a real voice matter more once an AI is answering for you?
Because your customers now meet you in more doorways than they used to. There is the website you already pay for, the Google listing, the Facebook page, a reply to a review, a text. If each one sounds like a different stranger, people feel it before they can say why. People search for answers, not websites, and a website stays invisible until a neighbor, Google or an AI points at it. When they do arrive, the voice is what tells them they have found a person.
That is the case for doing this once and doing it well. The same voice that answers the furnace email writes the Facebook post and the line for your St Louis Near Me Directory listing. If you can post to Google or Facebook, you can post here too. Same effort, and it adds up.
Want the plain-English version with a person beside you? The in-person AI workshops for St. Louis small-business owners that St Louis Near Me Directory runs are hands-on and plain English, with no jargon and nothing to buy.
If you’re the shop that wants more customers than word of mouth alone brings, remember who is on that couch. She is asking an AI who to phone. List your business on St Louis Near Me Directory and let your name be in the answer.
Frequently asked questions
How do I add custom instructions to ChatGPT?
On the web or desktop app, open Settings, choose Personalization, make sure Enable customization is on, and type your instructions into the Custom Instructions field. On the iPhone and Android apps the menu is called Customize ChatGPT. Updates to custom instructions settings are applied immediately across all chats, including existing conversations.
Can I use ChatGPT for my business?
Yes, with care. Individual plans can use your content for training unless you switch that off in Data controls, while OpenAI says Business, Enterprise and Edu are not used to train its models by default. Keep customer details out, check every fact it writes, and never let it invent credentials. Our longer guide covers the rest.
What shouldn’t you tell ChatGPT?
Passwords, card numbers, and any customer’s name, phone number, address or account details. On Free, Plus and Pro, OpenAI may use your conversations to train its models unless you turn off Improve the model for everyone. For something sensitive, use a temporary chat, and still leave the identifying details out.
Can I create my own AI like ChatGPT?
Not a new custom GPT on a personal account. OpenAI says new GPT creation and publishing are not available on Free, Go, Plus or Pro, and that it plans to retire custom GPTs in favor of plugins. What a small business can do today is build a Project with samples and instructions, which gets you a voice-trained helper inside ChatGPT.
Do project instructions replace my custom instructions?
Inside that project, yes. OpenAI says project instructions apply only within the project and override your global custom instructions. Your account-wide custom instructions still apply to all your other chats. So keep your never-invent-a-fact rule in both places, and let each project add the tone for its own kind of writing.
