Before You Prompt, Define the Job

Better prompts begin before you open the AI tool. Write down the job, the person who will use the result, the source material, the limits, and the check. The model has less to guess, and you have something concrete to evaluate.

People often treat a weak answer as a wording problem. They add a role, another instruction, three adjectives, and a demand to “think step by step.” Sometimes the answer improves. Often the prompt is carrying a job nobody defined.

Imagine asking an AI to “write a report about last week’s maintenance.” The request leaves almost everything open. Which equipment? Which records? Who is reading it? Does the reader need a summary, a decision, or a list of actions? What must never be left out? The model will fill those gaps with plausible choices. Plausible is not the same as right.

Start With the Result

Name the thing you expect to have at the end. “Help me with the meeting” is a wish. “Turn these notes into a one-page handoff for the Monday shift” describes an output. You can inspect a handoff. You can ask whether it includes the open issue, its owner, and the next action.

The result should connect to something a person needs to do. A summary may help someone catch up. A comparison may support a purchase. A draft may give an editor a first version. A classification may decide which queue receives a request. The action tells you what detail matters.

Name the Reader

The same facts need a different shape for a technician, manager, customer, or new employee. Naming the reader changes the vocabulary, depth, order, and assumptions.

“Explain this fault to a plant manager who needs to decide whether production can continue” is clearer than “make it easy to understand.” Easy for whom? A manager may need consequences and options. A technician may need measurements and the order of checks. A customer may need a plain explanation and the next update time.

If more than one group will use the output, decide whether one document can serve them or whether you need two versions. Asking one answer to be brief for leaders, detailed for specialists, reassuring for customers, and complete for auditors usually creates a compromise nobody likes.

Choose the Source Material

AI can write smoothly around missing information. That makes source discipline important. State what the answer may use: a manual, meeting transcript, spreadsheet, policy, set of emails, or collection of support tickets.

Give sources a priority when they can disagree. A current approved procedure may override an old training document. A signed contract may override notes from a sales call. A measured value may matter more than a general description. Write the order down.

Ask the model to separate sourced facts from assumptions. For high-consequence work, require a reference beside each claim or a short list showing where each conclusion came from. If the source does not contain the answer, “not found” is more useful than a polished guess.

Set the Boundaries

Constraints define the working area. They can cover length, format, reading level, deadline, confidential data, forbidden assumptions, required terminology, or decisions the AI must leave to a person.

The boundary matters most where the output can affect someone else. An AI may draft a reply while a person approves it. It may group incidents but not assign blame. It may compare options but not authorize a purchase. The prompt should make that stopping point visible.

Do not ask the model to hide uncertainty. Ask it to mark uncertainty in a consistent way. A short “Needs confirmation” section gives the reviewer a place to start.

Design the Review Before the Answer

Choose the check while you are still calm and before a polished response makes you want to believe it. The check should match the job.

For a handoff, verify that each open item has an owner and next action. For a comparison, check every price and product limit against a current source. For a draft email, check the names, dates, promises, and tone. For extracted data, compare a sample against the original records.

“Looks professional” is a reaction. “Contains all five required fields and no unsupported numbers” is a check. A useful test can fail and tell you why.

The Five-Part Worksheet

1. Job: What must be produced?
2. Reader: Who will use it, and for what action?
3. Sources: Which material counts, and which source wins?
4. Limits: What may the AI decide, and where must it stop?
5. Check: Which facts, fields, or conditions determine whether the result is usable?

Here is a finished example:

Notice that the example contains no theatrical role such as “You are the world’s best maintenance expert.” It gives the model the decisions already made and leaves it less room to invent the rest.

When the Answer Is Still Weak

Use the worksheet to diagnose the failure. If the structure is wrong, clarify the result. If the language misses the audience, describe the reader. If facts are wrong, improve the source and citation rule. If the answer crosses a line, strengthen the boundary. If you cannot tell whether it is good, repair the check.

A second prompt may still be necessary. Now you know what it needs to fix. You are correcting a defined job instead of adding random instructions until the output looks better.

This approach also makes tool comparisons fairer. Give two models the same job, sources, limits, and check. You can compare the work instead of comparing two lucky first attempts.

Further reading: OpenAI’s prompting guide and Anthropic’s prompting overview both emphasize clear instructions, context, and success criteria. Product recommendations change, so use the current documentation.

Next guideA Five-Question Check for Any AI Idea →