One of the stranger features of generative AI is that a small change in how we ask a question can produce a substantial change in the answer. Ask for “some ideas for improving customer retention” and you may receive a collection of familiar suggestions. Explain what kind of business you operate, where customers are leaving, what has already been tried, what constraints matter and what evidence you have, and the same model can produce something considerably more useful.
This has created an industry around “prompt engineering”, complete with formulas, templates and supposedly magical combinations of words that promise to extract better answers from AI. Some techniques are useful, particularly where prompts need to be repeated reliably across an organisation, but the underlying principle is considerably less exotic. A good prompt is mostly a well-specified problem.
The quality of an AI response depends partly on what the model can do, but also on what it has been given to work with. If the task is ambiguous, important information is missing and the criteria for a good answer are unclear, the model has to fill those gaps itself. Sometimes it will make reasonable assumptions. Sometimes it will make assumptions that are entirely plausible but wrong. Better prompting reduces the amount of consequential guessing the model has to do.
Start with the task, not the format
A common approach to prompting begins with the desired output: write me a report, create ten ideas, produce a strategy, summarise this document or give me a table. These instructions specify the shape of the answer without necessarily specifying the problem that needs to be solved.
Consider a manager asking:
Develop a communications strategy for our new service.
There is nothing technically wrong with the request, but almost every important variable remains unspecified. Who is the service for? What behaviour is the communication intended to change? What does the audience already know? Is the objective awareness, acquisition, retention or trust? What channels are available? What is the budget? What evidence exists about previous campaigns? What would success look like?
Without this information, the model will generally produce something that resembles a communications strategy. That distinction matters. Producing a document with the headings “Objectives”, “Target Audience”, “Key Messages”, “Channels” and “Measurement” is not the same as developing a strategy appropriate to a particular organisation.
A stronger prompt begins by defining the decision or problem before requesting the artefact:
We are launching a subscription service for small professional firms. Trial registrations are strong, but conversion from the 30-day trial to paid subscriptions is only 12 per cent. Most users who convert do so after using the reporting feature at least three times. We have a modest marketing budget and an existing email list. Develop a communications strategy focused specifically on improving trial-to-paid conversion rather than generating additional registrations.
The model now has a problem to work on rather than a document to imitate.
Context is not decoration
People often add background information to prompts as though they are briefing another person, but context has a more direct function in a language model because it changes the information against which subsequent output is generated.
If you ask:
What are the risks of introducing AI into this process?
the model must infer what kinds of risks matter.
If you instead specify that the process involves identifiable customer data, produces recommendations reviewed by qualified staff, operates under Australian privacy requirements and will initially be used only for internal decision support, you have materially changed the problem.
Good context therefore consists of information that should change the answer.
This is also a useful discipline for deciding what not to include. Longer prompts are not automatically better prompts. Pages of irrelevant organisational history can dilute the information that actually matters, while a few carefully selected facts can radically improve the response. The objective is not maximum context, but relevant context.
Tell the model what you know, and what you do not
One of the most useful prompting habits is to distinguish evidence from assumptions.
Suppose you are considering whether a fall in sales reflects a pricing problem. You could ask:
Why have our sales fallen?
The model can readily generate plausible explanations, including price, competition, seasonality, product quality, changes in customer preferences or weaker marketing. What it cannot determine from the question is which explanation is actually supported.
A better approach is to provide the available evidence and make the uncertainty explicit:
Sales have fallen 14 per cent over the past six months. Website traffic is broadly unchanged, average selling prices increased 8 per cent four months ago, conversion has declined from 4.2 to 3.5 per cent, and we do not currently have competitor pricing data. Identify the explanations supported by these observations, alternative explanations that remain possible, and the additional evidence required to distinguish between them.
This changes the task from storytelling to analysis.
The difference is particularly important in business settings because generative models are very capable of supplying missing causal explanations. A coherent explanation can sound like a finding even when it is only a hypothesis. Asking the model to separate what the evidence supports from what would require further information creates a much stronger analytical process.
Give it criteria, not just instructions
“Make this better” is a difficult instruction because better has no defined meaning.
Better could mean shorter, clearer, more technically rigorous, more persuasive, more accessible, more formal or more appropriate for a particular audience. Unless the criterion is specified, the model has to infer it.
The same problem appears in more consequential tasks. Asking AI to identify the “best option” gives it enormous discretion over what best means. Cost, quality, speed, risk, accessibility and strategic fit may point towards different choices.
Instead of:
Which of these proposals is best?
try:
Compare these proposals against implementation cost, expected time to deployment, dependence on external suppliers, data-security requirements and likely operational benefit. Explain where the evidence is insufficient to make a comparison.
The model now has an evaluative structure rather than an invitation to manufacture one.
This is particularly valuable when using AI as part of organisational decision-making. Criteria make assumptions visible. They allow a person to disagree not only with the conclusion but with the basis on which the conclusion was reached.
Examples can be more useful than adjectives
People frequently try to control AI output through adjectives: professional, engaging, sophisticated, concise, strategic, natural.
These instructions can help, but they leave considerable room for interpretation. “Professional” does not mean exactly the same thing in an academic paper, board paper, client proposal and LinkedIn post.
Where the distinction matters, an example can communicate the requirement more precisely than a collection of descriptive terms.
If you want a particular structure, provide a short example of the structure. If you are classifying customer enquiries into categories, provide examples of each category. If you need the model to distinguish a genuine operational risk from a general concern, show it one or two examples of each.
This is sometimes called few-shot prompting, but the practical principle is straightforward: demonstrate the distinction you want the model to make.
Examples are particularly valuable when the desired output depends on tacit conventions that are difficult to describe. An organisation may have its own definition of a qualified lead, serious incident, high-priority complaint or acceptable evidence. Giving the model examples helps translate those institutional conventions into the immediate task.
Ask for analysis before recommendations
Generative AI makes recommendations very easily, sometimes too easily.
Ask how an organisation should use AI and it can immediately propose chatbots, automated reporting, predictive analytics, personalised marketing and knowledge management. The list may sound sensible while having little connection to the organisation’s actual problems.
A stronger sequence begins with diagnosis.
Rather than:
Recommend five AI use cases for our finance team.
ask:
Based on the workflow information below, identify where staff spend substantial time on repetitive information processing, where delays occur because information has to be manually transferred between systems, where decisions depend on unstructured documents, and where errors are currently expensive. Do not recommend technologies yet.
Once those problems have been identified, the next instruction can ask which are suitable candidates for AI, conventional automation, process redesign or no intervention.
This separation matters because otherwise the requested solution can distort the diagnosis. If you ask for AI opportunities, the model has every incentive to find them. If you first ask it to identify operational problems, you create the possibility that AI is not the appropriate answer.
Ask the model to challenge the premise
Questions contain assumptions, and language models are often very good at accepting them.
How should we use AI to reduce employee turnover?
contains an implicit assumption that AI is an appropriate intervention and perhaps another that the drivers of turnover are amenable to technological intervention.
A better prompt might say:
We are considering whether AI could contribute to reducing employee turnover. Before proposing any AI intervention, identify the principal assumptions behind that proposition, what evidence would be required to test them, and circumstances in which AI would be unlikely to address the underlying problem.
This is a particularly useful technique when the person writing the prompt has a preferred solution. Asking the model to challenge the framing can expose assumptions that would otherwise become embedded in the response.
You can go further by asking what would have to be true for a proposal to work, what evidence would cause the model to reject its initial conclusion, or what a credible critic would identify as its weakest assumption. These questions use the model less as an answer generator and more as a mechanism for testing an argument.
Separate facts, inference and speculation
One of the simplest ways to improve research and analytical prompts is to ask the model to distinguish different epistemic categories.
For example:
For each major conclusion, distinguish between information directly supported by the supplied documents, an inference reasonably drawn from that information, and a hypothesis that would require additional evidence.
This is useful because fluent prose can otherwise collapse those categories. A fact can lead to an inference, which leads to an assumption, which is then presented three paragraphs later as though it were part of the original evidence.
The same technique can be used when working with uncertain quantitative information:
Do not estimate missing figures unless I explicitly ask you to. Where a value is unavailable, identify the missing information and explain how it could be calculated.
That single instruction can be considerably more valuable than telling a model to “be accurate”.
Give permission to say that information is missing
Many prompts inadvertently create pressure for completion.
If you ask for ten findings, the model is encouraged to produce ten findings. If the evidence supports only six, the remaining four still have to come from somewhere. If you request a complete market assessment from incomplete data, the model will often try to produce something resembling a complete market assessment.
The better instruction is often conditional:
Identify up to ten findings that are supported by the evidence. Do not add additional findings simply to reach ten.
Or:
Assess each criterion only where the supplied information is sufficient. Mark other criteria as requiring additional evidence.
This sounds like a small adjustment, but it changes the incentives built into the task. Completion is no longer treated as more important than evidentiary restraint.
Use AI iteratively
The idea of the perfect prompt is probably one of the least useful ideas to come out of the early generative AI period.
For complex work, trying to specify everything correctly in a single enormous prompt is often less effective than working through the problem in stages. The first interaction can establish the problem, the second interrogate assumptions, the third identify missing information, the fourth analyse the evidence and the fifth produce the final artefact.
This resembles good analytical work more generally. We rarely begin a serious research project by writing the final report. We define the question, inspect the evidence, test interpretations, identify gaps and revise our thinking.
AI works particularly well when the conversation is treated in this way because each stage can create useful context for the next. An early discussion about objectives and constraints can subsequently inform the analysis, while identified uncertainties can be addressed before recommendations are generated.
The important distinction is between iteration and repeated requests to “try again”. Asking for another answer without changing the information, criteria or reasoning process may simply produce another variation on the same weakness. Useful iteration changes the problem the model is being asked to solve.
Sometimes the best prompt is a question back to you
There is another simple technique that is particularly useful when you are unsure what information matters:
Before answering, ask me the questions you need answered to do this properly.
This reverses the usual relationship. Rather than requiring the person to anticipate every relevant variable, the model can help identify what is missing.
For a communications strategy, it might ask about the audience, objective, channels, budget and previous performance. For a software project, it might ask about users, existing systems, security requirements and deployment constraints. For research, it might ask about the population, available data, identification strategy and intended use of the findings.
This is often a better use of AI than immediately requesting an answer because formulating the problem is itself part of the work.
A practical structure
There is no universal formula for a good prompt, and rigid templates can become counterproductive. For substantive work, however, five elements are worth considering:
Task: What are you actually asking the model to do?
Context: What information would materially change its answer?
Evidence: What information should it rely on, and what is currently unknown?
Criteria: Against what standards should alternatives or outputs be assessed?
Constraints: What must it not assume, change or invent?
A useful prompt might therefore read:
Assess whether introducing an AI-assisted triage system would be appropriate for the customer support process described below. We receive approximately 4,000 enquiries each month, around 60 per cent of which fall into six recurring categories. Staff currently classify enquiries manually before assigning them to specialist teams, and misclassification creates an average delay of 1.8 days. Use the supplied workflow and service data as the evidence base. Assess potential benefits against implementation complexity, privacy, classification error and staff oversight requirements. Distinguish findings supported by the data from assumptions, identify information that is missing, and do not assume AI is preferable to conventional rules-based automation.
There is nothing magical about that prompt. It is simply a reasonably well-specified analytical task.
And that is ultimately the most useful way to think about prompting. The objective is not to learn a secret language for communicating with machines. Contemporary models are increasingly capable of understanding ordinary language, and elaborate prompt formulas will matter less as models become better at interpreting intent.
What will continue to matter is the quality of the question.
A vague problem expressed through an elaborate prompt remains a vague problem. A poorly defined objective does not become clear because the instruction assigns the model a role or tells it to “think step by step”. Conversely, a clearly specified problem, supported by relevant context, explicit evidence and sensible criteria gives the model much better material from which to produce a useful response.
The transferable skill is therefore not prompt engineering in the narrow sense. It is problem formulation: knowing what you are trying to establish, what information matters, what assumptions should be questioned and what would constitute a good answer.
Those were valuable skills before generative AI. The technology has simply made the quality of our questions much more visible.