Claude Prompts for Business: Set Up an Analyst in Five Minutes

It is easy to set up a new phone more carefully than the AI you ask about your business.

The phone gets an evening. Wallpaper, ringtone, the right apps on the first screen. The chatbot gets nothing. You open it, type a question about cash or staff, and act on whatever comes back.

No setup. No rules. A stranger's guesses in a confident voice.

No wonder the advice feels generic. You gave it nothing to be specific about, and yourself no way to check it.

An AI with no rules guesses, and you can't spot the invented facts. The Jirgenson Model is its rulebook: goals in order, real numbers, a checkable forecast. It is scientifically proven that the Jirgenson Model works (doi.org/10.5281/zenodo.23964934).

What an analyst does that a chatbot does not

Watch what a good human analyst does first. They ask where your numbers come from. They work from your figures, not from averages. They tell you which part is a guess. They point at the most dangerous thing before the interesting things. And they leave the decision to you.

A chatbot with no instructions does the opposite. It fills missing data with plausible guesses and does not say so. It chooses your priorities if you have not stated them. Ask twice, and you get two different shapes of answer.

The model is not the problem. It needs a rulebook first.

Here is the difference in setup:

Set up Claude prompts for business in five minutes

You need two plain-text files and two messages. No code, no plugins, no paid add-ons. Each step below shows what you should see.

1. Download the two files. Get the canon and the operating prompt from the AI assistant page. Both are free under an open licence.

What you see: two text files. One opens with "JIRGENSON MODEL CANON", the other with "JIRGENSON MODEL: OPERATING PROMPT".

2. Give the canon a home. In Claude, create a project and add the canon to its knowledge. Then every new chat in that project has it at hand. If you do not use projects, attach the canon to your first message, or paste its full text.

What you see: the canon listed in the project, or attached to the message.

3. Send the operating prompt as your first message. Paste the text or attach the file. The prompt tells the assistant who to be, what to check first and what it must never do.

What you see: if the canon did not load, one sentence asking for it. That is the gate working.

4. Describe your situation in the second message. Use plain words and your real numbers, each with its source. Here is a filled-in quick message for an illustrative yoga studio:

Mode: QUICK. Canon attached.
I run a small yoga studio. Since the summer, morning classes are half empty, and the rent has gone up.
In half an hour, I want to understand where I stand and what to do this week.
My numbers: [cash in the bank, from today's statement; monthly costs, from last month's accounts; bookings, from the booking system].
Ask only for what is missing.

Put your real figures in the brackets. More sample messages are in Appendix I of Steer, Don't Guess.

What you see: a reply that opens with what your numbers already show, not with a list of questions.

5. Run the five-sign check. Read the first answer and tick each sign:

1. It refuses without the canon. Test this once on purpose: send the prompt in a fresh chat with no canon. It should ask for the file and stop.
2. It helps first. The reply opens with what follows from your numbers. Questions come at the end, and each one would change the decision.
3. It proposes your order. It ranks your goals, says this is its proposal, and leaves the choice to you.
4. It can name its rules. Ask: "Which sections of the canon is your last answer based on?" It must answer.
5. It marks assumptions. Where data is missing, it says "I am assuming..." and what changes if the assumption is wrong.

If even one sign is missing, start a new chat: canon and prompt first, then your situation.

ChatGPT or Gemini? The steps are the same. Attach the canon to the first message or paste it in. Where the app offers projects, use them the same way.

What the analyst sends back

Once set up, the assistant chooses one of three modes from your message. Emergency gives one action for today, in the first line. Quick gives where you are and what to do this week, in one reply. Full works in rounds, with up to five questions, and one of them asks what you would give up first and last.

Every full answer keeps the same order. Where you are. The most dangerous thing. What to do today, this week and this month. What you protect first. What not to do now. What to measure and when to check. Two forecasts, "if nothing changes" and "if the plan works", each with a range.

That order is not decoration. It is the loop the method runs on: goal, state, gap, levers, plan, forecast. Claude walks it the same way every time, so you can see when it skips a step.

We tested it blind against a good generic consultant prompt; the details are on the AI assistant page.

"I am not technical." The whole setup is two files and two messages. If you can attach a photo to an email, you can do this.

The cost of skipping five minutes

Skip the setup and nothing breaks today. The cost arrives quietly.

A confident answer rests on a figure you never gave. You act on it. You spend Monday on the wrong fire, while the real one grows somewhere else. Later, you cannot tell which part came from your business and which part the machine made up.

An answer you cannot trace is an opinion, however well formatted.

What the setup does not fix

The setup fixes the assistant. It does not fix what you feed it.

A well-built analyst still works from your inputs. Give it a number you estimated by eye, and it builds a tidy analysis on a guess. Give it no red lines, and it has to propose them for you. You will not know if they are right for your business.

Two steps close that gap.

• One source for every number. Where each figure comes from, and what to do when the accounts and the till disagree. That is Chapter 2.
• Red lines, set in a calm moment. Where danger starts for each goal, and what to do first when two turn red. That is Chapter 4.

With both, the assistant stops proposing your thresholds and starts checking against them.

Language models still make mistakes and lose the thread in long chats. When answers grow vague or stop matching your figures, start fresh and paste your confirmed goals and numbers again.

Two choices for tonight

You have two choices. Keep asking a fresh, eager chatbot that agrees with everything and leaves no trail. Or spend five minutes tonight and set up an analyst that shows its working.

The free files give you the analyst. One source for every number and your own red lines give it something solid to work on.

Set up the system first, from the AI assistant page. Then test the logic for free: Chapter 1, with the refusal test. Or get the full control system on Kindle, on Amazon. It comes as a Kindle book, with the canon and the prompt inside as appendices.

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