A lot of bad prompts are really badly defined jobs. Before changing the wording for the fifth time, write down what you actually need.
Take a request such as “Write a report about last week’s maintenance.” It sounds clear until you try to use the answer. Which equipment belongs in the report? Who is reading it? Does that person need a summary or a decision? Are old work orders relevant? The AI has no way to know unless you tell it.
When those choices are missing, the model makes them for you. The result may sound convincing and still be useless.
Describe the finished result
Start with the thing you want to have when the work is done. “Help me prepare for the meeting” leaves too much open. “Turn these notes into a one-page handoff for Monday’s shift supervisor” gives you something you can inspect.
Then ask what the reader should be able to do with it. Catch up before a meeting? Approve a purchase? Assign open tasks? The answer changes what belongs in the document.
“The reader should be able to ______ without having to ______.” Fill in both blanks. You will often find the real job hiding there.
Say who will use it
“Make it easy to understand” is not very helpful. Easy for a technician is different from easy for a customer or a plant manager.
A manager deciding whether production can continue probably needs the consequence, the available options, and the next decision. A technician may need measurements and the order of checks. Same fault. Different document.
If two groups need very different things, make two versions. One answer rarely manages to be short for management, detailed for technicians, reassuring for customers, and complete for an audit.
Choose what the AI may use
Tell the tool where the facts come from. That could be a transcript, a spreadsheet, a manual, a set of emails, or the notes from a technician.
If the sources can disagree, decide which one wins. The current approved procedure should probably outrank an old training file. A signed contract carries more weight than somebody’s meeting notes. Write this into the instruction instead of hoping the model works it out.
For work where a wrong fact matters, ask the AI to mark anything it cannot find. “Not in the supplied material” is a perfectly good answer. A confident guess is not.
Decide where the AI stops
Some limits are about the format: length, required headings, terminology, or reading level. Others are about authority.
The AI might draft a customer reply, but a person approves the promise. It can group incidents, but it should not decide who caused them. It can compare prices without approving the purchase. Put that line in writing.
Uncertainty needs a place as well. I like a short “Needs confirmation” section. It tells the reviewer where to look first.
Know how you will check it
Do this before you see the answer. Polished writing is surprisingly good at lowering your guard.
The check depends on the job. A handoff needs every open item, owner, and next action. A price comparison needs current prices. A draft email needs the right names, dates, and promises. Extracted data should be compared with a sample of the original records.
“Looks professional” is an opinion. “All open work orders are included and every status matches the export” is a test.
A five-part worksheet
- Job: What must be produced?
- Reader: Who will use it, and for what action?
- Sources: Which material counts, and which source wins?
- Limits: What may the AI decide, and where must it stop?
- Check: Which facts, fields, or conditions determine whether the result is usable?
Here is what that looks like when the blanks are filled in:
Create a one-page maintenance handoff for the incoming shift supervisor. Use only the attached technician notes and current work-order export. List each unresolved issue, affected equipment, current status, owner, and next action. Do not estimate repair times or assign causes that are not in the sources. Put missing information under “Needs confirmation.” The result passes when every open work order appears once and its status matches the export.
There is no “You are the world’s best maintenance expert” at the top. It is not needed. The useful part is that the decisions have already been made.
If the answer is still poor
Go back to the worksheet. A messy structure usually points to an unclear result. The wrong tone often means the reader was not described well enough. Wrong facts call for better source material or a stricter rule about using it.
You may still need a second prompt. At least now you know what you are correcting. That is quicker than piling on random instructions until one version happens to look good.
The same worksheet is useful when comparing two tools. Give both the same job, material, limits, and check. Then compare the output rather than two lucky first attempts.
Next guideA Five-Question Check for Any AI Idea →