You tried an AI tool to answer a customer or write a follow-up, and what came back sounded fine and was wrong. It offered a slot you do not have, quoted a warranty you do not offer, or wrote three paragraphs where you needed two lines. You are not a prompt engineer and do not want to be. You want it to sound like your business and get the facts right.
Here is the short answer. A good business prompt does five things: it says what job the AI is doing, gives it the business facts it must use, describes the exact shape of the answer, lists what it must never do, and says when to stop and hand the job to a person. Most bad answers come from missing facts, not bad wording, so keep those facts written down in one place.
Key Takeaways
- Brief the AI like a new hire: the job, the facts, what done looks like, the hard limits, and when to come get you.
- The most common failure is missing business facts. The AI cannot know your cancellation policy or your warranty terms unless you hand them over.
- Describe the output shape in plain words: how long, what tone, what to include, what to leave out.
- Test any prompt you plan to reuse against five real past cases before you trust it with a customer.
- Keep your policies, prices, and "how we do it" notes in one written place, so every prompt uses the same facts.
What are the parts of a good business prompt?
Think of a prompt as a brief for a sharp temp who has never set foot in your shop. If they could not do the job without calling you, the AI cannot either.
- The job. One or two sentences: who the AI is writing as and what it is producing. "You are replying on behalf of Bright Cuts Salon to a client who wants to reschedule." Not "be a helpful assistant."
- The business facts it must use. Your actual policy, prices, hours, or warranty terms. Paste them in. This is the part people skip, and it matters most.
- The output shape. Length, tone, format, and what must be in it. "Under 80 words. Friendly, not gushing. End with one clear question."
- What it must never do. The lines it cannot cross. "Never promise a time that is not in the openings list. Never offer a discount. Never mention other clients."
- When to hand it to a person. The cases where the AI should stop and flag it instead of guessing. "If the client is upset, mentions a reaction or injury, or asks for a refund, do not reply. Write HAND TO OWNER and one line on why."
That fifth part is the one owners forget. For help deciding which jobs should always go to a person, see AI vs human: when to automate a business task.
Example 1: replying to a reschedule request at a salon
Before:
Write a reply to this client who wants to reschedule her appointment.
The AI will write something polite and invent the rest: "any time that works for you," no mention of your late-change fee, her stylist on a day that stylist is off.
After:
You are writing a text message reply for Bright Cuts Salon to a client who wants to move her appointment.
Facts to use: Rescheduling policy: [paste your actual policy, including any late-change fee]. Openings this week with her stylist, Dana: Thursday 2:00, Thursday 4:30, Saturday 10:00.
Shape: Under 60 words. Warm and short, like a text from the front desk. Offer the openings listed, nothing else. End by asking which one she wants.
Never: offer a time not in the list, waive the fee, or book anything. You are only drafting.
Hand to a person if: she is upset, asks to switch stylists, or her message mentions a problem with her last visit. In that case write HAND TO OWNER and one line on why.
The second version is longer, and that is fine. Every added line closes off a wrong answer.
Example 2: an estimate follow-up for a roofing job
Before:
Write a follow-up email for a roofing estimate we sent last week.
You will get a generic "just checking in!" email, possibly with a made-up discount to "close the deal."
After:
You are writing a follow-up email from Ridgeline Roofing to a homeowner who received an estimate and has not replied.
Facts to use: Estimate sent [date] for a full tear-off and replacement, [square footage], [shingle line]. The estimate is valid for [your real validity period]. Our workmanship warranty is [paste your exact terms]. We are currently scheduling installs about [your real lead time] out. The homeowner mentioned being worried about the old decking.
Shape: Under 120 words. Plain and respectful, no pressure. Address the decking concern using our decking terms from the estimate. Remind them when the estimate expires. Offer one next step: reply with a question or pick a start week.
Never: offer a discount, change the price, guarantee a start date, or say anything about insurance coverage.
Hand to a person if: the homeowner has replied with a counteroffer, mentioned an insurance claim, or said they went with someone else.
The brackets are facts only you have. Leave them out and the AI fills them in with something plausible, which is how a roofer ends up "promising" a warranty they do not offer.
Example 3: summarizing the week's reviews
Before:
Summarize these reviews.
You get a cheerful paragraph saying customers are "generally satisfied," which you cannot act on.
After:
You are summarizing this week's online reviews for the owner of a two-location plumbing shop. The owner reads this on Monday morning and has five minutes.
Facts to use: Our locations are North and Downtown. Our techs are [names]. Our after-hours fee is [amount], and it is disclosed on the phone before dispatch. Reviews are pasted below.
Shape: Three short sections. 1) Anything that needs a reply or a call today, with the reviewer's name and the one-line reason. 2) Patterns: anything mentioned by two or more reviewers, good or bad. 3) Tech names that came up, with the gist. No overall star average, no fluff.
Never: draft public replies, guess at what happened on a job, or call a review fake.
Hand to a person if: a review mentions property damage, a safety issue, or a legal threat. Put those at the very top, marked URGENT.
Knowing your fee and your tech names, it can connect several complaints about "surprise after-hours charges" to a disclosure problem instead of reporting "some pricing concerns."
How do you test a prompt before you trust it?
A prompt that works on the first example can still fail on the fourth. Before you use one with customers, run it against five real past cases.
- Pull five real past cases. Pick five actual examples of this job from your texts, emails, or reviews. Include an easy one, one where the customer was annoyed, and one you handled differently from the usual.
- Write down what a good answer would have been. A sentence or two each. You already know.
- Run the prompt on each one. Paste the case in, read what comes back.
- Check each answer against four questions. Did it use the right facts? Did it invent anything? Was it the right length and tone? Did it hand off the cases it should have?
- Fix the prompt, not the answer. If it invented a time slot, add a "never" line or the missing fact. Then run all five again, because a fix for one case can break another.
When all five come back right, save the prompt. When one later goes wrong on a real customer, add that case to your test set. If you are still choosing which job to start with, the first 5 processes to automate with AI is a good shortlist.
Why do most business prompts fail?
Wording gets the blame, but the usual cause is that the AI did not have the facts. It did not know your cancellation window, your service area, or that you stopped offering a service in the spring. So it filled the gap with something that sounds normal for a business like yours.
The trouble is where those facts live. In many small businesses they are spread across the owner's head, an old price sheet, a text thread with the office manager, and a website page that is out of date. So every prompt gets the facts retyped from memory. One person remembers the new fee and another does not, and the AI repeats whichever version it was handed.
So the most useful thing you can do for your AI output is not about prompting. Write your business facts down in one place, and keep them current. Policies, prices, service area, warranty terms, "how we handle a complaint" notes. Every prompt copies from that one source, and when a price changes you change it once. For more on how this goes wrong, see common AI automation mistakes and how to avoid them.
Where Mycel fits
Mycel is built around that last point. Your team keeps the business's written knowledge (policies, prices, how-we-do-it notes, spreadsheets) in one shared vault of plain folders. Mycel's chat assistant answers from that vault, and the Martello apps you connect, such as Relay for customer messages or TradesBackbone for field-service work, read from the folders you choose to share when their AI drafts a reply. Anything that would change something inside another app waits for your approval.
Mycel is not a prompt builder or an automation tool, and it is not a CRM. It is the place your facts live so the AI you use has the right ones. You can read how Mycel works or check pricing; there is a free tier.
Frequently Asked Questions
How long should a business AI prompt be?
As long as it takes to cover the job, the facts, the output shape, the hard limits, and the hand-off rules. For a simple reply that is often a short paragraph plus your pasted policy. Cut any line that does not prevent a wrong answer.
Do I need to learn prompt engineering to use AI in my business?
No. If you can write instructions for a new employee, you can write a good business prompt. What matters is knowing your policies and being specific about a good answer, which you already do.
Why does AI make things up about my business?
Because it was not given the real facts, so it filled the gap with what sounds typical. Paste your actual policy, prices, or terms into the prompt, and ask it to flag anything those facts do not cover.
Should I let AI send replies to customers automatically?
Start with drafts a person reads before sending. Once a prompt has handled your test cases and a few weeks of real ones without surprises, decide whether low-risk replies, like appointment confirmations, can go out on their own. Money, complaints, and safety stay with a person.
Can I reuse the same prompt in different AI tools?
Usually, yes. Plain instructions and real facts travel well between tools. Run your five test cases again when you switch, since tools differ in how closely they follow length and format instructions.
Your next step this week
Pick one job you handle every week, like reschedule requests or estimate follow-ups. Write the five-part prompt, then test it on five real past cases from your inbox. Note every fact you had to type from memory, and start a single document with those facts in it. Every prompt after this one gets easier.