In this article
The important distinction is between starting an experiment and trusting its result. You can start simply while remaining demanding about accuracy, confidentiality and usefulness.
Start your AI prompting with something you can evaluate
Choose a task you understand well enough to inspect. An experienced operations manager can spot a misleading handover note more readily than a plausible answer about an unfamiliar legal issue.
For a first experiment, use invented details or approved, non-sensitive material. Try turning a rough explanation into a clearer outline, preparing questions for a meeting or comparing possible structures for a presentation.
Ethan Mollick's good-enough prompting essay encourages learning through practical dialogue. Treat that as an invitation to experiment, rather than permission to send unreviewed work.
Before opening the tool, finish this sentence: “A useful answer would help me _____.” If the blank is only “be productive,” narrow it until you can recognise a satisfactory output.
Write a brief that makes the job visible
A useful starting brief answers four questions: what should be produced, who will use it, what information supports it and what must be avoided?
Here is a hypothetical workplace example:
Help me prepare a handover note for a colleague covering tomorrow's store opening. Use only the fictional notes below. Organise the result into confirmed actions, unresolved questions and items requiring manager approval. Do not invent owners, deadlines or completed checks.
That brief does more than assign a role. It defines a job and protects the gaps that might otherwise disappear inside polished prose.
For a different task, keep the same logic while changing the substance. A customer explanation needs the customer's concern and the approved position. A meeting agenda needs a decision, participants and relevant constraints.
Leave out unnecessary personal information. More context is useful only when you are allowed to share it and it helps the task.
Improve the answer by naming the defect
The first response is evidence about your instructions and the tool's limitations. Inspect it before asking for a more impressive version.
“Make this better” gives little direction. More actionable feedback might be:
- “The opening hides the decision. Put the requested action first.”
- “You turned an unresolved question into a promise. Restore the uncertainty.”
- “The audience knows the background. Remove the first two paragraphs.”
- “Separate what the notes explicitly say from your suggested next steps.”
Change the problem you can see. If the information is missing, obtain it from the right source; do not keep asking the model to sound more confident.
OpenAI's prompting guidance describes explicit instructions and evaluation as useful practices. A personal experiment can apply the same principle with a simple checklist.
Judge quality before judging speed
Use a small review grid rather than relying on whether the output sounds professional.
| Check | Question to ask |
|---|---|
| Accuracy | Can each factual statement be traced to approved information? |
| Completeness | Did the answer preserve the important exception or unresolved issue? |
| Audience fit | Can the intended reader understand what to do next? |
| Effort | How much checking and correction did the output require? |
Suppose the hypothetical handover looks excellent but silently assigns the opening inspection to the wrong person. That is a failed result, even if the formatting took seconds.
Record the correction. It may reveal a recurring instruction worth adding to your brief: “List an owner only when the notes explicitly name one.”
Know when simple prompting stops being enough
A one-off draft and a repeated business workflow need different controls. If colleagues will reuse the prompt, build a few representative examples and check whether it behaves consistently.
Include an awkward example: incomplete notes, contradictory dates or a request outside the intended task. Decide what the tool should do in those cases. Asking for clarification can be a better outcome than producing an answer.
Avoid using this beginner exercise to make employment, medical, legal or financial decisions. Where consequences are significant, the responsible professional and organisational process must determine the review required.
A worked revision: from tidy to trustworthy
Suppose the fictional opening notes say: “Display delivery may arrive before opening. Ask the duty manager whether the entrance needs moving. Inspection owner not yet agreed.” A weak AI draft might turn that into: “The opening supervisor will move the entrance display and complete the inspection before opening.”
The problem is not tone. The draft invented a delivery certainty, an authorised change and an owner. Correcting only the grammar would leave the most important errors untouched.
Give targeted feedback: “The delivery is uncertain, the display change requires a decision and no inspection owner is agreed. Rewrite these as unresolved items. Preserve the distinction between an action and a question.”
Then compare the revision with the original notes. A satisfactory result would leave the delivery tentative, name the duty manager as the decision contact and flag the missing inspection owner. It would not silently solve the operational problem.
This example also reveals a limit on prompting. The tool can organise the gap, but someone in the real workflow must resolve it. When you understand that distinction, a modest prompt can be useful without being mistaken for a complete solution.
Keep a small record of what you learned
After an experiment, write down the task, the useful instruction, the main failure and whether you would try the approach again. A handful of meaningful observations is more valuable than a folder of impressive prompts you have never used.
For a next step, choose one non-sensitive task and allow two revisions. Stop when you have either a checked result or a clear reason the tool is unsuitable.
Explore the PiOfAI hub for the broader topic, then connect the experiment to a personal AI workflow or a repeatable AI procedure. The aim is dependable work, beginning with an understandable brief.
AI-assisted researched guidance from PeakBelt. Hypothetical examples are labelled in the text. See our editorial approach.