How to write a good AI prompt: the science, engineering, and human judgment behind a good instruction

AI can process large amounts of information, but it does not necessarily know the context a person takes for granted. We know what happened in yesterday's meeting, what the client wants, which version of a document is current, why a particular metric matters, and what decision we are trying to make. A model can only work with what is part of its available context or with information it can access through the appropriate tools.

This difference seems simple, but it explains much of the frustration people experience when they begin using AI. They ask something very general, receive an equally general answer, and conclude that the tool ‘didn't understand.’ Often, the problem is not that the system ignored the instruction, but that it never received enough information to correctly interpret what the user had in mind.

That is why human planning remains the first step. Before asking how we should write the prompt, we should ask what problem we are trying to solve, what information we have available, which data is reliable, and what a useful response would look like. That preliminary organization is not a task for AI. It is part of the intellectual work of the person directing the process.

At Alterno, we see this principle as fundamental. AI can expand capacity, accelerate processes, and reduce mechanical work, but direction still starts with a person.

A prompt can be a single question, but for professional tasks, it is usually more useful to think of it as a collection of information that guides the model. The instruction is only one part. Documents, previous conversations, data, examples, rules, constraints, and prior results can also form part of the context.

If we ask ‘summarize this report,’ we are leaving it to the model to decide what it considers important. If, instead, we explain that the summary will be used by management, that it should focus on budget and performance changes, that it should compare against the previous quarter, and that any missing data should be flagged rather than estimated, we have completely changed the task.

We are not simply giving it more words. We are giving it more structure.

That is one of the central ideas behind a good prompt: reducing the number of important things the model has to guess.

Large language models process information using units called tokens. A token can represent a complete word, part of a word, punctuation marks, or other linguistic units. Instructions, provided documents, previous conversation, and other available data take up space within what we call the model's context window.

Based on that context, the system generates a response by sequentially estimating what information is most likely and appropriate to produce. Modern models based on Transformer architectures use attention mechanisms to establish relationships between different parts of the context and determine which elements may be relevant to the task.

This helps explain why a small difference in the information provided can significantly change a response. The instruction ‘analyze this campaign’ opens up many possibilities. Should we analyze creative, spending, audience, media, conversion, or profitability? If we specify that we want to compare CPA, conversion rate, and ROAS against the previous period, the problem begins to become much more clearly defined.

The technology may be sophisticated, but the principle for the user is surprisingly human: if you want a more specific answer, you need to frame a more specific problem.

Today's models can handle increasingly large amounts of information. We can attach lengthy documents, reports, spreadsheets, manuals, policies, presentations, and other materials. That capability is extraordinarily useful, but it can also create a misconception: that the more information we send, the better the result will automatically be.

Not necessarily. A relevant hundred-page file can be invaluable. Twenty partially related documents, duplicate versions, and outdated material can introduce noise. The technical ability to process a lot of information does not eliminate the human need to decide which information truly matters.

In a sense, this is very similar to research. A researcher does not improve a study simply by accumulating sources. They have to select, classify, contextualize, and establish priorities among them.

Something similar happens with AI. A good practice can be to explain which document is the official source, which contains background information, and which should be used only for comparison. If two files contradict each other, we can also establish which takes priority. The model receives more than documents: it receives a structure for interpreting them.

This introduces a less technical and more organizational dimension. Many companies operate thanks to an enormous amount of knowledge that has never been formally documented. One person knows how to prepare a particular report because they have been doing it for ten years. Another knows that certain clients require a different approach. Someone knows which figures need to be reviewed before sending a presentation. Another person knows which version of a process actually works, even if the manual says otherwise.

That knowledge is, to a large extent, tacit knowledge. It lives in people.

Using AI seriously often requires turning some of that knowledge into explicit information. We have to explain criteria, document rules, establish priorities, and describe exceptions. For a system to help us with a process, we first need to understand that process well enough to explain it.

That can be a positive consequence of AI adoption within organizations. Prompt engineering does not just help the machine work better. It can push us to better define how we work.

Suppose we finish an hour-long meeting with a client and have the full transcript. We can write: ‘Summarize this meeting.’ We will probably receive an acceptable summary. But the model will have to decide for itself what summarizing means.

Now imagine we write: ‘I have attached the transcript of a meeting with a client. I need an executive summary for the internal team. Identify the decisions that were confirmed, the issues left unresolved, the tasks mentioned, and any dates discussed. Do not assign owners unless they are explicitly identified in the conversation, and do not invent dates. The summary should allow someone who did not attend to understand what was decided and what needs to happen next.’

The difference is not in using special words. The difference is that someone first thought about the purpose of the summary. That is prompt engineering in a professional context.

Agencies, consultants, developers, and business teams have used briefs for decades for a simple reason: it is difficult to produce good work when the person doing it has to guess what the requester expects. The same principle applies to AI.

A good prompt usually clarifies the task's objective, provides essential context, identifies relevant sources, and establishes what result we need. It can also explain which information should be excluded, how to handle missing data, and the most useful format for the response.

This does not mean every request has to become a five-page document. The complexity of the prompt should reflect the complexity of the problem. Asking for the capital of France does not require additional context. Analyzing three years of advertising results probably does.

As prompt engineering became more popular, enormous templates full of instructions, roles, steps, and commands also began circulating. Some can be useful, but there is a risk of confusing complexity with quality.

A prompt can contain 2,000 words and still be ambiguous. It can also contain contradictory instructions. We might simultaneously ask for a response to be extremely comprehensive and extremely brief, to leave out no details and to stay within two paragraphs. The number of instructions does not resolve those contradictions.

More recent models have also reduced the need for some techniques previously used to guide less capable systems. The current trend in documentation from leading providers favors clear instructions, relevant context, and a precise definition of the task.

The right question is not ‘How do I make the prompt longer?’ It is ‘What does the model need to know that it does not know yet?’

One of the most useful capabilities of today's platforms is working directly with documents. This significantly changes how AI is used in professional settings.

Imagine we want to understand a company's benefits policy. We could try to explain the policy to the model from memory. Or we could provide the official handbook and ask it to use only that document to answer. The second option is much more reliable.

The same applies to reports, proposals, contracts, research, manuals, spreadsheets, market studies, or technical documentation. When the information exists, providing the source is usually better than trying to reconstruct it within the prompt.

But even then, human judgment remains important. We have to determine which document is current, which contains the correct information, and what the system should do if it cannot find an answer.

The opposite mistake also exists: attaching twenty documents and writing only ‘analyze them.’ The AI now has information, but it may still not understand what problem we want to solve.

We can improve the task considerably by explaining that certain files contain results, others contain the budget, and another contains the project's original goals. We can request a specific comparison and establish which metrics should be used.

Context and instruction work together. Documents primarily answer the question ‘With what information?’ The prompt should answer ‘For what purpose?’

Another important technique is providing examples. In prompt engineering, this practice is usually called few-shot prompting. Instead of only explaining a rule, we show a few cases that correctly represent the pattern we want the model to follow.

This is especially useful in repetitive processes. Imagine a company that receives hundreds of messages and needs to classify them as ‘Sales,’ ‘Technical Support,’ or ‘Billing.’ We can explain what each category means, but we can also provide several real examples that have already been classified. The model then has concrete references for how to apply the criteria.

The example does not replace the rule. It complements it. For operational tasks, classification, information extraction, or data transformation, this combination can considerably improve consistency.

AI models can produce extremely convincing responses. That is precisely one of the reasons we need to design tasks carefully. If a figure, date, or source is missing, we do not necessarily want the system to try to fill the gap.

A simple instruction could be: ‘If the information does not appear in the provided documents, state that clearly. Do not estimate or invent the data.’

These types of rules are particularly important in research, business analysis, documentation, finance, and other contexts where accuracy matters more than fluency. The ability to produce language should not be confused with a guarantee of truth. That is why a good prompt also defines the limits of the response.

Sometimes we try to use AI as though all the work has to happen within a single instruction: ‘Read these twelve documents, identify problems, compare them, prepare recommendations, develop a plan, and make me a presentation.’ The system can try. But there is another way to work.

First, we can ask it to identify the relevant data. Then review that extraction. Next, ask for a comparison. Later, develop hypotheses and finally organize recommendations.

This approach has two advantages. It reduces the complexity of each step and allows a person to validate the process as it moves forward. That matters because human oversight should not happen only at the end. It can be part of the entire execution.

The concept of the ‘perfect prompt’ can also create the expectation that we should write a flawless instruction on the first try. In practice, professional prompting looks much more like a process of testing and refinement.

We ask for something. We review the result. We identify where it misinterpreted the task. We add context. We remove instructions that created noise. We try again.

Over time, we can even turn a prompt that works consistently into a reusable template for a recurring task. The engineering lies precisely in that iteration. Not in finding a magic phrase.

One of the most useful ways to begin a complex task is not to ask for the final result immediately. We can provide the objective and write: ‘Before you begin, review the information I provided and identify what important data is missing that you need to complete this task correctly.’

That can reveal problems in our own brief. Perhaps we never defined the period we want to analyze. Maybe two documents have different versions of the same figure. Perhaps we did not specify which metric determines success.

In those cases, AI is helping structure the question better before attempting to answer it. And a good question is often the beginning of a better answer.

At Alterno, we make an important distinction between using AI to expand capacity and using it as an automatic substitute for professional talent.

There are processes where technology can deliver tremendous value: organizing information, summarizing documentation, comparing data, accelerating research, classifying content, detecting inconsistencies, automating repetitive tasks, or structuring information so a person can work with it more easily. That does not mean every activity should be automated.

Our philosophy continues to place human talent at the center of creative processes. Designers, art directors, photographers, videographers, editors, copywriters, and other professionals bring experience, sensitivity, cultural context, and judgment that we do not reduce simply to the ability to generate an output.

The existence of a generative tool does not automatically make using it the best decision. The question should be where AI truly adds value.

Human creativity does not arise solely from combining information. It also comes from experience, observation, culture, intuition, memory, sensitivity, interpretation, and the ability to understand other people.

AI can help organize material related to a creative problem. It can summarize research or identify patterns within a study. It can reduce some of the operational work surrounding a project.

But direction requires intent. Someone has to understand what a brand wants to communicate, why it matters, how an audience might interpret it, and what consequences the decision will have. The tool can contribute. The judgment remains human.

One principle runs through this entire topic: a clear prompt is often the result of clear thinking.

If we do not know the objective, it will be difficult to explain it to AI. If we do not know which information is relevant, we will probably send too much or too little context. If we do not know how to evaluate a good response, we will not know when the result is good enough either.

In that sense, prompt engineering is not only a technology skill. It is also a communication skill. An analytical skill. A synthesis skill. An organizational skill. And a decision-making skill.

People who develop these capabilities will probably be better able to use AI systems, regardless of which platform or model is available five years from now.

Another important factor is that a prompt does not exist in a vacuum. Models evolve. Their reasoning capabilities, tools, context windows, and behavior change. A technique that was particularly useful with one generation may become unnecessary with another. Even the same prompt can produce somewhat different responses across models.

That is why, in important business applications, writing an instruction and assuming it will always work is not enough. It is worth testing it with real examples and checking whether it continues to produce acceptable results when the inputs change. Here we begin to enter another part of AI systems engineering: evaluation.

The quality of a prompt should not be judged by how sophisticated it looks. It should be judged by how well it works.

If we are using AI to extract information from invoices, we can compare the extracted data with the original documents. If it classifies leads, we can review a sample and calculate how many classifications were correct. If it summarizes reports, we can check whether it preserves the essential figures and conclusions.

In more advanced AI implementations, these tests are known as evals: systematic evaluations of the model's behavior against known cases. For an everyday user, the concept can be much simpler: do not just ask whether the response ‘sounds good.’ Ask whether it is correct, complete, and useful.

It is important to maintain realistic expectations. A good prompt can considerably improve a result, but it does not guarantee perfection. The model may interpret something incorrectly. It may omit information. It may use the wrong source. It may produce a statement that seems plausible but is unsupported.

That is why the idea that everything depends on the prompt would also be an exaggeration. The model, available tools, sources, data quality, and nature of the task all matter. And human review still matters. Professional responsibility does not disappear because a response was produced by a machine.

We are probably entering a stage where knowing how to use AI will increasingly be less of a specialized skill and more of a general workplace competency. But the advantage will not necessarily come from memorizing hundreds of prompts. It will come from knowing how to structure problems. Select information. Recognize reliable sources. Define criteria. Detect inconsistencies. Ask good questions. Evaluate answers. And use the right tool when it truly adds value.

That also explains why organizations with good processes can benefit particularly from AI. A team that already documents procedures, keeps information organized, and clearly defines its objectives has an excellent foundation for working with models.

AI can accelerate processes. But first, there has to be a process worth accelerating.

Perhaps the best answer is that the perfect prompt does not exist. There is a prompt that is clear enough for a specific task. There is relevant context. There are reliable sources. There are instructions that reduce ambiguity. There is a definition of what we expect to receive. There is a review process. And there is a person capable of determining whether the result truly has value.

That is why the science behind a good prompt begins with technology, but does not end there. It also involves how we, as people, turn knowledge into language, how we structure problems, and how we communicate our expectations.

AI can process an extraordinary amount of information. It can analyze. It can organize. It can accelerate. But someone has to give it direction. The quality of the result depends largely on the quality of the context we provide. And the quality of that context still begins with us.

What is prompt engineering? It is the practice of designing and improving the instructions, context, sources, and criteria we provide to an AI system to increase the likelihood of obtaining correct, relevant, and consistent results. It is not just about writing a question; it can include documents, examples, rules, background information, and a clear definition of the expected result.

How do you write a good AI prompt? A good prompt begins by defining the objective. Then it provides the context the model needs to understand the situation and specifies what information it should use, what constraints exist, and how we want to receive the result. For complex tasks, it can also be useful to divide the process into several stages and review each one before continuing.

Does a longer prompt mean a better response? Not necessarily. A longer prompt can provide useful context, but it can also introduce irrelevant information or contradictory instructions. Clarity, relevance, and structure are more important than word count.

Does uploading documents help produce better responses? It can help significantly when the documents contain information the model needs to complete the task. A manual, report, PDF, or spreadsheet can provide specific context that would be difficult to explain manually. However, it is still important to indicate what to look for within those documents and what the objective of the analysis is.

What is an AI context window? It is the amount of information a model can consider during an interaction. It can include the prompt, previous messages, documents, and other inputs. A larger window allows the model to work with more information, but it does not eliminate the need to select and organize context properly.

Can a good prompt prevent AI errors? It can reduce ambiguity and significantly improve the quality of a response, but it does not completely eliminate the possibility of errors. That is why important results should be verified and kept under human oversight.