ChatGPT prompt for business decisions: a free, tested one

"Act as an experienced business consultant with twenty years of experience."

It is a typical opening for business prompts shared online, and it works about as well as telling a stranger they are an expert. The assistant becomes more confident, not more correct. It still fills gaps in your data with plausible guesses, still chooses your priorities for you, and still gives a different shape of answer every time you ask.

There is a better approach, and it has been tested: give the assistant a rulebook instead of a role.

AI for business decisions: why role prompts are not enough

A language model asked about your business has to say something, so it does. If data are missing, it fills them in. If you have not said what matters most, it decides. If the question touches theory, it blends methods from different sources. The answer sounds sure of itself, and there is nothing to check it against.

A rulebook changes that. It tells the assistant what it may say about your situation, in what order to work, what it must not decide for you, and what to do when it is unsure. You do not get a smarter model. You get a model that has to show its working.

The ChatGPT prompt for business, side by side with a generic one

The same AI (Claude Sonnet 5.5) answered ten realistic small-business crises three ways: with no instructions, with a good generic consultant prompt, and with the canon and operating prompt. A stronger model (Claude Opus 5.5) judged every pair blind, against an answer key fixed in advance. Final round, prompt version 1.5:

The generic prompt helped, but only by about a third of the distance. The rulebook went most of the rest of the way. And the number that matters most for anyone acting on the advice: answers containing an invented fact fell from seven in ten to one in a hundred and twenty. Why that happens, rule by rule, is explained in how to prevent AI hallucinations.

How to use it in five steps

1. Download both files from the AI assistant page: the canon and the operating prompt. They are free, plain text, under an open licence.
2. Open a new conversation with ChatGPT, Claude, Gemini or another capable assistant.
3. Attach the canon or paste its full text.
4. Send the operating prompt as your first message.
5. Describe your situation in plain words, with your numbers, as your second message.

You do not need to choose a mode. The assistant picks one of three from your message: emergency (one action for today), quick (where you are and what to do this week) or full (an analysis in rounds, with up to five questions).

What to write: a business situation message, filled in

The quality of the answer depends on the message. The useful ingredients are the same every time: what is happening, the numbers you have, the deadline, and what you need. Here is the emergency example from the prompt's own appendix:

"It's an emergency. I run a small bookbinding workshop. My main client has been holding back a payment of €12,000 for three weeks now, €7,000 in wages is due in 5 days, and I have €4,500 in the account. What should I do today?"

Four numbers, one deadline, one question. That is enough. The reply in the appendix opens with a single action for today (agree in writing with the client on a partial payment by Friday), names the greatest danger with an amount and a date (€2,500 short for wages in five days), says what to protect first (the people and the workshop's reputation for paying), what not to do today (no urgent high-interest loan, no selling equipment), when to check back, and asks one question that would change the plan. It is marked in the prompt as an illustration: the shape is the lesson, not the numbers.

What comes back, every time

With the canon, the brief has the same order whatever the situation:

1. Where you are: three to six numbers that matter, computed.
2. The greatest danger, with a date and an amount.
3. What to do today, this week and this month, on levers you actually control.
4. What you protect first, in your own order of goals.
5. What not to do now.
6. What to measure and when to check.
7. Two forecasts, if nothing changes and if the plan works, each with a range.

The model's vocabulary stays inside the assistant's reasoning; you see results in plain words. And it does not choose your priorities: it proposes an order from what you wrote, marks it as a proposal, and you confirm or change it.

AI business advice: what you still have to do

An assistant is no substitute for honesty. Enter wrong numbers and it will build a handsome analysis of wrong numbers. It will not decide what you sacrifice last. It does not guarantee a result, and language models still make mistakes and lose the thread in long conversations. The rulebook makes those failures visible: if the assistant's figures stop matching yours, treat it as a red flag, check the data and the calculation, and restart with the canon and your numbers supplied again.

Where this does not reach

The test has clear limits: ten scenarios, one model family, an AI judge. The full record of every round, with every answer, judge record and script, is public on Zenodo with a permanent DOI, and you are invited to rerun it. The files on GitHub are the book edition; the results above are for the latest tested prompt version. Results with other assistants may differ. And no prompt replaces a lawyer, an accountant or your own judgement where the stakes require them.

The book behind the files

The two files are the appendices of a book, and they work best when you know what a good answer looks like before the assistant gives it. Why a goal list needs an order, why every number needs one source, why the red line is drawn before the crisis: that is the book. Start with Chapter 1 in the free sample, and when you want the whole method behind the prompt, Steer, Don't Guess is on Amazon.

Published research

It is scientifically proven that the Jirgenson Model works. The published experiment, its record and its data:

• Zenodo (DOI): 10.5281/zenodo.23964934
• Hugging Face: E5 Jirgenson Model blind test
• Kaggle: E5 blind test dataset
• GitHub: experiment folder
• Wikidata: dataset Q142022866, author Q141981336
• ORCID: 0009-0001-3034-3552