How to write prompts for ChatGPT and Claude. Also tell the AI what not to do

The owner of a wholesale business types "Sort out the reports" into ChatGPT and attaches a file. What comes back is a long, polite answer about how reports can be organised in general, and none of it ends up in any document. The problem is rarely the model. More often it is the request, and how to write prompts that give a usable result is something you can learn without a course.

Why "Sort out the reports" does not work

A request like that leaves every decision to the model. Which reports, sorted by what, in what shape, and what happens to the result afterwards. The model does not know your business, so it guesses, and the guess comes out vague.

"Make it look nice" has the same problem. To you, nice means something specific, maybe short paragraphs and a table at the top. The model does not know that, so it makes things nice in its own way.

A good request reads like a task for a new colleague who is sharp but on their first day. Anything you leave out, they will have to guess.

If you are not sure your request is complete, add "If anything is missing, ask me first" at the end. The model will ask a few questions, and you will see straight away what you left out.

How to write prompts that come back ready to use

Every request should answer these questions:

1. What exactly to do. One task, named with a verb. Count, compare, draft a reply.
2. From what. Which file, which column, which period. If the source is attached, say what it is called.
3. In what form. A table, a list, an email or a short paragraph. With which columns, or how long.
4. Where it will go. An email to your accountant or a slide for a meeting. That decides the tone and the length.
5. What not to do. For example "do not delete anything and do not send anything", "do not change any figures", "do not add anything that is not in the file".

The last point matters more and more, because AI assistants can now work with files and email when you give them access. A clear limit costs you one sentence.

A full example

The data below is made up for this article. Say you export a file called sales-july.csv from your sales software, with the columns date, salesperson, client and net_amount. Client names have already been replaced with "Client A", "Client B" and so on.

Calculate July sales separately for each salesperson. Take the data from the attached file sales-july.csv, from the columns "salesperson" and "net_amount". Show the result as a table with the columns salesperson, number of sales and total net, sorted from the highest total. I will paste the table into an email to the management, so no introduction and no comments. Do not change or delete any data in the file. If a row has no salesperson, do not guess, list that row separately under the table.

Each sentence answers one of the questions from the list. The last one is there because a model that does not know what to do with a gap may simply make something up.

The same request works in Claude. Before you send the table on, check at least one salesperson's total by hand.

Ask for changes in the same conversation. "Add a column with the number of clients" works better than a new request from scratch, because the model still has the file and everything you have agreed so far.

Habits that save you a round of fixes

For bigger tasks, ask for options first. "Before you start, suggest a few ways to lay out this report and ask me which one I want." A plan is easier to correct than a finished result.

You can also dictate. When people speak, they tend to give more detail than when they type, and detail is exactly what the model needs. The microphone on your phone keyboard is enough, and some AI apps have their own dictation too.

Do not paste your clients' personal data. Write "Client A" and "Client B" instead of names, and leave addresses out unless they really matter. For a sales table the model does not need to know who the client is. Whether and on what terms company data can go into these tools is worth checking with a lawyer, especially with GDPR in mind. We wrote about one ChatGPT setting worth checking before you paste anything.

If you type the same request every week, that is a sign the task could be automated for good. That is what we do in AI automation.

Start with one task you do every month. Write the request for it following the list above and save it in your notes, so next time you only swap the file.