ChatGPT Prompts for Business Owners, By Task

Most lists of chatgpt prompts for business ask the model for a number you don't have. "Compare these to industry benchmarks for [your industry]." "Competitor prices: [list if known]." Analyse the market size. One of the ranking pages, LivePlan's, tried the market-size prompt on three platforms and reports getting "confident, yet 100% made-up responses", and then keeps the prompt in the list anyway.

An owner's prompt is safe when three things hold. Every number in it is yours. The model does arithmetic, sorting, and drafting, and supplies no fact you didn't paste. And the output goes to a named checker before it's used, which is sometimes you, sometimes the person who does the task next, and sometimes the accountant. Six prompts below hold to that, one per function, pricing, cash, operations, hiring, customers, and the meeting with the professional. Each is shown in full, with a worked calculation for the pricing one so you can see what it should return.

Before any of them, there's a switch to check, and it's at the end.

What The Owner Supplies And What The Model May Not

The lists sort prompts by department. I'd sort them by who has to check the output, and I'd put that column before the prompt, because that decides what the model is allowed to contribute. A grouping of your own cancellation messages needs only your eyes. A job ad goes in front of strangers and, depending on where you are, the law. A tax question goes to the accountant, and the model's role is to make that meeting shorter.

The fourth column is where the results page goes wrong. Techpresso's Financial Health Dashboard prompt collects seven of your own metrics and then asks the model to compare them to benchmarks it was never given. The Layer3 Labs hiring prompt asks for "must-have skills, nice-to-have skills, and pay range" from a role title and a description of the business. Those prompts don't fail loudly. They come back fluent, formatted, and with a number in every field, and OpenAI's own page for GPT-6 Astra describes a model that "uses the context it has to fill in routine gaps". A benchmark is a routine gap.

Pricing, With The Arithmetic Written Out

Use only the numbers I give you. Do not add benchmarks, industry averages, competitor prices, or price elasticity figures. If a calculation needs a number I haven't given, write [NEED: what] and stop that branch.
Current price: [P]
Cost to deliver one unit: [C]
Fixed costs per month: [F]
Units sold per month, last three months: [N1], [N2], [N3]
Proposed price: [P2]
Show, with every line of arithmetic written out:
1. Break-even units per month at the current price and at the proposed price.
2. Monthly profit at the current price, using the average of the three months.
3. The units per month at the proposed price that give the same profit.
4. The share of current volume I could lose at the new price before profit drops below today's.
Round only in the last line of each section and say when you do. No recommendation.

This is what the pricing prompt on the results page should have been. Techpresso's version collects the same inputs and then asks for an evaluation "using value-based, cost-plus, and competitive pricing frameworks", which is three essays and no arithmetic. The version above returns four numbers you can check with a calculator, and "No recommendation" keeps the model from telling you what to charge, which is a decision that depends on things it doesn't know.

Here's the calculation with illustrative inputs, so you know what a correct answer looks like. Price 40, cost to deliver one unit 4, fixed costs 600 a month, 30 units a month for three months running, proposed price 48. None of those are anyone's real numbers.

Line seven is the one to keep. It says the price rise pays for itself as long as fewer than roughly one customer in five walks, and whether that's likely is the question the model can't answer and you can. I sell seven digital toolkits priced between $29 and $49, and for a digital product the cost line is close to zero, which makes the same arithmetic run almost entirely on fixed costs and volume. The structure doesn't change. The unit cost slot just gets a small number.

Check line two and line six yourself before acting. The model tends to get this arithmetic right when it's forced to show it and wrong more often when it isn't, which is a judgment from use rather than a measurement, and the prompt is built for the first case.

Cash, Thirteen Weeks, No Guessing

Build a 13-week cash table from the entries below, starting [DATE]. Columns: week, opening cash, cash in, cash out, closing cash. Use only the entries I list. Apply a recurring entry on the weeks I state and nowhere else. Do not estimate receivables, do not assume an invoice is paid on time, and do not add a category I didn't give. An entry with no date goes in a separate list called "undated" with its amount and is left out of the table. Mark the first week where closing cash goes below [MINIMUM] with the word FLOOR and show that week's arithmetic in full.
Opening cash on [DATE]: [AMOUNT]
Cash in, one per line as amount, date or weeks, and certain or expected:
[PASTE]
Cash out, one per line as amount, date or weeks:
[PASTE]

The Reddit thread at the top of the results page complains that business plans come back "too vague, overly optimistic", and the optimism lives in one assumption. Invoices get paid when they're due. This prompt refuses that assumption by name. If you mark an inflow "expected" it still goes in the table, because the point of the table is to see what happens if the expected money is late, and the way to see that is to delete the line and run it again.

The undated list is the second output, and in my judgment it tends to be longer than you'd like. Every item on it is a cash movement you know about and haven't pinned to a week, which is the definition of a surprise. FLOOR is a word rather than a colour so it survives being pasted into a spreadsheet.

Operations, The Procedure From One Walkthrough

Below are my notes from doing [TASK] once, written in the order I did it, including where I hesitated. Write a standard operating procedure from these notes and nothing else. Number the steps and keep my order. Where a step depends on something I didn't write down, such as a login, a file location, or a person to ask, insert [MISSING: what] instead of filling it in. Where I wrote that I hesitated or wasn't sure, add a line under that step that begins "Decision point" and restates my uncertainty in one sentence. Do not add steps, best practices, or safety notes I didn't mention. End with the list of [MISSING] items.
Notes:
[PASTE]

Every SOP prompt on the results page is a one-liner, "write a step-by-step standard operating procedure for [task]", and what comes back is a procedure for a business the model imagined. Yours has to come from a walkthrough. Do the task once with a notes file open and write down each thing you did, badly and in order, including the moment you stopped to wonder whether the invoice number goes in the subject line.

The [MISSING] markers are the reason to run this. They're the list of things in your head that aren't in the notes, and the person who does the task next is the checker, since they'll hit every gap the markers didn't catch. In my judgment the "Decision point" lines are worth more than the steps. A procedure that hides where the judgment calls are is a procedure that gets followed wrongly with confidence.

Hiring, The Description With The Pay Left Blank

Write a job description from the task list below. Use only these tasks. Group them into day-to-day, weekly, and occasional, keeping my wording. Do not add responsibilities, qualifications, years of experience, degrees, or personality traits I didn't list. Leave the pay as [PAY RANGE, owner to fill] and do not suggest a figure. Do not mention age, health, family status, nationality, or a physical requirement unless I listed one as essential to a task. Plain sentences. No "rockstar", no "ninja", no "we're a family". Then, separately, list the qualifications a person would need to do these tasks, each marked "inferred", so I can accept or delete them one by one.
Role title: [TITLE]
The business, in one sentence: [SENTENCE]
Tasks:
[PASTE]

The pay range is blank on purpose. A model-suggested range is a benchmark by another name, and it'll be a plausible one, which is worse than an obviously wrong one because you might publish it. The protected-characteristics line is a filter, and I'm careful to call it that rather than legal advice. It catches phrases the model reaches for when it's writing in the register of a thousand job ads, and what's actually prohibited depends on where you're hiring, which is a question for the last prompt on this page.

The inferred qualifications go in a separate list because that's the part you're most likely to accept without reading. Mark them, read them, and delete the ones that describe the previous person rather than the tasks. If you want the candidate's side of the same document, the resume post covers what a job description looks like from underneath.

Customers, The Cancellation Pile

Below are [N] messages from customers who cancelled, complained, or asked for a refund. Group them by the cause the customer states, not the cause you infer. Name each group in the customer's own words, give the count, and quote one message per group in full. Messages that state no cause go in a group called "no cause stated" with a count. Do not add a group with zero messages. Do not turn counts into percentages unless the total is [N], and then show the division. Do not suggest fixes. Order the groups by count, largest first.
Messages:
[PASTE]

The lists call this churn analysis and ask for root causes. The word "infer" is the problem. Given twenty cancellation notes the model tends to produce a tidy set of causes with a percentage next to each, and the percentages are what get repeated in the next meeting. Grouping by the stated cause and quoting one message in full keeps the output honest and short, and my expectation is that "no cause stated" ends up the biggest group, which is itself the finding. Fixes are banned here because a fix proposed by a model that's read twenty complaints is a fix for the average complaint, and no single customer wrote that one.

The Accountant, Questions Instead Of Answers

I'm meeting my accountant about the situation below. Do not give me tax, legal, or accounting advice, and do not state what any rule, rate, threshold, or deadline is. Instead, write the questions I should ask, in the order that makes the meeting shortest, each with one line on why it matters and the document or number I should bring to answer it. If a question depends on where the business is registered, say so and leave the jurisdiction to me. At most ten questions.
Situation:
[PASTE]
Country and business type: [PASTE]

This is the prompt I'd copy if you copy one. Every page on the results page has a prompt that asks the model what the rule is, and the model will tell you, in the confident register LivePlan describes, with a rate and a date attached. The honest use of the model in a regulated question is preparation. It turns a vague worry into ten specific questions with the paperwork listed next to each, and the accountant, who charges by the hour, gets a client who arrives organised.

The ordering clause matters more than it looks to me. Asked for questions, the model tends to list them by topic. Asked for the order that shortens the meeting, it puts the questions whose answers eliminate other questions first, which is how a good accountant would run it anyway.

The Switch To Check Before Pasting A Ledger

The US Chamber's page warns owners to "avoid uploading sensitive business data" because "confidentiality cannot be guaranteed". True, and it stops short of saying where the control is.

OpenAI's help page on how data is used, in its September 2026 snapshot, says that on the individual products "we may use your content to train our models", and that you turn that off under Settings, then Data Controls, by switching off "Improve the model for everyone". After that, the page says, new conversations aren't used for training. Temporary Chat is exempt regardless, and it also doesn't appear in history or touch memory. For ChatGPT Business, ChatGPT Enterprise, and the API, the same page says inputs and outputs aren't trained on by default. One caveat from the same page. If you give a thumbs up or down on a response, the whole conversation attached to that feedback may be used for training, opt-out or not.

So the cash prompt and the pricing prompt go in a Temporary Chat or an account with the switch off, and I wouldn't rate the answer. That's the whole policy.

Questions From The Results Page

What are the top 10 ChatGPT prompts, and the top 10 AI prompts? A top ten is a top ten of someone else's business. The slots that make any prompt work, role, task, context, format, constraints, and examples, are in chatgpt prompts, and the split between text, image, and video prompting that most top-ten lists ignore is in ai prompts.

How do you make ChatGPT suggest your business? There's no form. Nothing in the OpenAI pages I read for this article describes a way to submit a business for recommendation, and I'd be suspicious of anyone selling one. The model answers from what it has read or can browse, so the route is the one you already know. A page that answers the question your customer asks, on a site the model can reach. Whether that works for your category is something I can't verify, and I'd rather say so than invent a mechanism.

What are some famous prompts for ChatGPT? The famous ones are on the results page. SWOT analysis, competitive analysis, the 90-day plan, the pitch deck outline. They're famous because they produce a full page from a bracket, and the full page is the problem. Run any of them and count the numbers in the output that you didn't type in.

What are the best AI prompts for business, and what are some good business prompts? The six above, with the checker column. The best one for the money is the accountant prompt, since it replaces the prompt most likely to cost you something.

Is there a PDF, and are these free? Free, and they copy cleanly into a notes file. I'd keep the pricing table next to the pricing prompt so that the next time you run it, you have a worked answer to compare the shape against.

What about business ideas, and the five prompts to start a successful business? LivePlan's page is the honest one here. It admits the market-size numbers were made up and still lists the prompt. My own numbers are small, roughly $10,000 lifetime across everything I've shipped, and that's the size of business where an invented benchmark does the most damage, because nobody else is going to catch it. Ideas are cheap and the model has thousands. The prompts on this page are for after you've picked one.

What about marketing? Campaign work, the brief, and the claims register that keeps copy honest are in chatgpt prompts for marketing. And the personal side, the prompts for your own week rather than the company's functions, is in useful chatgpt prompts.

What I'd Do First

Run the pricing prompt with your real numbers tonight, in a Temporary Chat, and check line six against a calculator. If the arithmetic holds, you've got a prompt you can trust for the next price change. If it doesn't, you've learned something about the model before it cost you anything.

Then do the walkthrough for one task you're tired of explaining. The [MISSING] list is the first useful thing the model will have written for your business, because every item on it was in your head and nowhere else.