Trash in, trash outs: The quality of the response depends on the first prompt

There is an ancient principle in software engineering, so old that many programmers know it by its acronym: garbage in, garbage out. If the input is garbage, the output will be garbage. It was true for the mainframes of the 1960s, it’s true for the spreadsheet you open every morning, and it’s true for today’s AI models. Except that with AI, the principle is called “trash in, trash out,” and it has more subtle consequences. From many perspectives, to one decision, yours: to arrive at an informed decision, the starting point matters just as much as the system processing it.

The Founding Principle

A generative AI model produces responses based on a statistical distribution of plausible completions. When the prompt is clear, the model narrows the distribution to relevant completions, and the resulting response is useful. When the prompt is vague, the distribution remains broad, and the model selects a generic completion that looks like an answer but doesn’t address your question.

This isn’t a flaw in the model. It’s a direct consequence of how it works. A generative model works on what it receives. If what it receives is imprecise, it works well on something imprecise. The result is a response that seems authoritative and well-formulated, but that doesn’t really match your question because your question wasn’t actually the one written in the prompt.

The practical consequence is counterintuitive. The most underrated lever in productivity with AI isn’t “finding the best model.” It is “writing the initial prompt better.” An average model with an excellent prompt beats an excellent model with a mediocre prompt in the vast majority of real-world use cases.

The Three Pitfalls

Three recurring pitfalls alone account for almost the entire problem of prompts that don’t work.

The first is the vague prompt. “Help me with marketing.” “Give me advice on this decision.” “What should I do?” The model has no context, no constraints, no defined task. It has to fill in the gaps by inventing assumptions about what you mean. Its assumptions aren’t yours. The response is generic by design, because the question never had enough structure to generate a specific answer.

The second is the ambiguous prompt. “Write me something brief but comprehensive on topic X, in a professional yet somewhat creative tone, for a general but experienced audience.” Each clause contradicts the other. The model must choose a single interpretation from many possible ones, and whatever choice it makes will betray at least half of your instructions. The response isn’t wrong: it’s one of the possible answers compatible with the instructions, but probably not the one you had in mind.

The third is the overloaded prompt. A massive request that muddles ten things together, with pages of context, scattered constraints, and mixed formats. The model, even when it has the technical capacity to handle a long input, tends to lose focus. The final parts of the prompt receive less weight. The instructions in the middle get lost. The result is a response that covers everything superficially and nothing in depth.

These three issues explain the vast majority of the “the AI didn’t give me what I wanted” complaints heard in meetings. The problem, almost always, wasn’t the answer. It was the question.

The anatomy of a prompt that works

A well-constructed prompt isn’t a masterpiece of writing. It’s a precise set of minimal information. Six elements, to keep in mind as a checklist.

The first element is the target role. Who is supposed to answer? An analyst? A consultant? A creative? Specifying the role isn’t a formality: it changes how the model interprets the question. The same prompt addressed to “a financial analyst” or “a brand strategist” produces structurally different answers, because the two roles bring different default contexts.

The second element is the context of use. Who is the answer for? In what situation will it be read? Is it an internal memo or an external presentation? Is it a personal decision or a topic to bring to a committee? The context of use determines the register, the depth, and the type of evidence needed.

The third element is the specific task. What exactly must the model produce? “Write,” “analyze,” “list,” “evaluate,” “critique.” Different verbs trigger different behaviors. “Write a memo” is not the same as “analyze a problem.” Being precise about the initial verb is one of the most effective levers.

The fourth element is explicit constraints. Time, length, tone, things to include, things to avoid. Constraints seem limiting, and in human writing they often are. For an AI model, they are the opposite: the more explicit the constraints, the more useful the response. A prompt without constraints is an invitation to vagueness.

The fifth element is the response format. List? Table? Numbered paragraphs? A single summary line? Specifying the format in advance avoids the most common problem: getting a response that’s correct in content but wrong in form, and having to resubmit to reformat it.

The sixth element, optional but powerful, is the success criterion. How do you measure a good response? “A useful response is one that allows me to X.” Stating this guides the model far more effectively than any qualitative adjective. “Be precise” means nothing. “A useful response is one my CFO can use in a meeting without further work” means a lot.

These six elements don’t all need to be stated every time. They should be kept as a mental framework. When an answer turns out poorly, the problem is almost always in one of the six points that had been left implicit.

Why the Problem Persists

If the diagnosis is so clear, why hasn’t the problem of poorly constructed prompts already been solved? Three reasons converge.

The first is that formulating a good prompt is a cognitive task, and humans tend to avoid it whenever possible. It’s faster to throw out a vague question and see what comes up. Except that what comes up requires ten rounds of refinement, and overall, you end up spending more time than you would have spent formulating the question well from the start.

The second is that "prompt engineering" isn’t taught in school, in the workplace, or as a general practice. It’s learned through trial and error, slowly, and many people stick to naive phrasing even after years of intensive AI use. The "garbage in, garbage out" principle is well-known as a slogan but rarely practiced.

The third, and most interesting, is that many AI models have learned to “make up for” vague prompts by producing answers that seem good even when they aren’t. You ask for little, receive something plausible, and are satisfied on the surface. The model has become very good at hiding the problem without solving it. This reinforces the habit of poor phrasing.

The Structural Solution

The individual solution is to learn how to construct better prompts. The structural solution is to build systems that process the quality of the prompt before it reaches the model that must respond.

This shifts the leverage point. Instead of asking every user to become an expert in prompt engineering, the system inserts a layer of intelligence between the user and the generative model. That layer questions the user to understand what is truly needed, reconstructs intent, context, and constraints, and produces the structured prompt that feeds the next step in the workflow.

In practice: instead of an empty text field asking the user “what do you want,” a structured system asks “tell me what you’re trying to do,” then follows up with “what context are you in,” then “what is the most important constraint,” then “what do you need certainty about.” The user isn’t writing a prompt; they’re having a natural conversation. The prompt emerges from that conversation, built by a product component, not improvised by a nervous user who has five minutes.

This structural solution solves two problems at once. The first is the quality of the prompt: it reaches the model in a form the model can work with effectively. The second is the user experience: the user isn’t forced to become an expert in prompt engineering to get useful answers. The system does the work that would otherwise require cognitive effort unrelated to the problem at hand.

What Changes for Decision-Makers

If you use AI to inform important decisions, the “garbage in, garbage out” principle ceases to be merely academic and becomes operational. A decision made based on an AI response to a vague prompt is a fragile decision, because that response was one of a thousand possibilities, and you haven’t seen the others. Refining the quality of the prompt is not a tooling activity; it is an exercise in decision-making rigor.

Refining the prompt doesn’t just mean “formulating the question better.” It also means stating the constraints you often keep implicit. What resources do you actually have? What is the non-negotiable success criterion? What is the most important risk to avoid? These are things we often don’t write down because “we know them.” The AI model doesn’t know them. Writing them down forces you to make them explicit even to yourself, and this alone improves the quality of the decision, regardless of the response you receive.

What Arena Does

Arena structurally addresses the “garbage in, garbage out” problem with a dedicated product component. At the start of the workflow, even before selecting a team, the user may be greeted by an interview agent that reconstructs intent, context, constraints, and success criteria, and generates the first structured prompt that will feed the workflow. It’s not a rigid wizard with mandatory fields: it’s a natural conversation.

From there on, the well-constructed prompt is read in parallel by many complementary specialists, each with a different focus. The responses arrive together; the user selects the most useful ones; the Orchestrators writes the final report and proposes the next prompt—already constructed but editable by the user. The refinement loop inherits the quality of the first prompt and amplifies it.

This means that the “garbage in, garbage out” principle is no longer the responsibility of the individual user. It becomes a structural property of the workflow. Transparency: everything remains visible, no black boxes—you see how your first prompt was structured, you can modify it, and you can understand why a certain formulation produces certain responses.

Change the way

Change the way you use AI. Change the way you make informed decisions.

Enter the Arena because the quality of the answer starts with the quality of the question, and the quality of the question cannot remain the user’s problem: compare, choose, explore, decide, starting from a well-constructed initial prompt created together with the system.