Users are seeing low-quality content, wrong answers, hallucinated facts, and other unbelievable output from AI. The output is so prevalent on the internet that Merriam-Webster made slop the Word of the Year in 2025. The masses do not like it. In fact, only about 10% of Americans are more excited than concerned about AI. But the slop isn’t necessarily an AI problem. It is often an instructions problem.
Consider the case of a boss who gives an employee a task but no clear instructions. To get the task done, the employee will fill in any gaps with their own judgment. It is possible for that to work sometimes, but usually it doesn’t. That is analogous to how people are largely approaching AI. They give poorly constructed task instructions and receive poor results. With AI, these instructions are called prompts. And most need improvement.
Quality prompts define a task, the inputs, and the expected output. Setting up prompts in a way that the AI better understands leads to dramatically improved output. It sends the AI in the right direction, but it cannot solve all the problems associated with AI. I highlight remediating those problems in a related article.
The following prompt is one of several from my Culture Research Engine project:
Prompt Structure
A decent prompt structure for better output typically has four layers.
Role
Think of the role as the person you would call for assistance in the task you want to perform. The example prompt states, “You are a music supervisor.” The reasoning is that a music supervisor and a researcher in anthropology have different languages and frames of reference in their specialties. Adding the role primes the model for these factors before it receives anything else. The role is rarely a determining factor in output quality, but it does edge the AI in the right direction.
Task
Unlike the role, the task is much more important. The task tells the AI what it must do. Keep it simple. The task should be one clear sentence with only enough detail required to decrease ambiguity. Any ambiguity will the foundation for everything downstream. In my example, “Based on the topic and gathered research, generate a Suno style prompt for an instrumental background track.” It is a straightforward task, nothing extra. The Culture Research Engine incorporates several prompts in its script, each doing a different. It gives cleaner results.
Constraints
Constraints guide the AI how to produce output. These can include format requirements, inclusions, exclusions, and other rules. The example constraints include a several square bracket annotations like [max 120 characters] or [comma-separated, all lowercase]. These tell the AI what form the output should take. When adding constraints, be explicit with what you want. Implicitness yields unreliable results. When adding rules, reinforced rewordings can help. In the example, “Instrumental only — never include vocals or lyrics tags” works better than “instrumental only.” Without constraints in place, the model will default to verbosity and excessive formatting, which requires manual cleanup later.
Data
The data is not only the raw material from which the AI will work, it is also a key component to reduce hallucinations. You know where the data comes from, so your confidence level can be a little higher, but not a guarantee. Data should always come last. The AI needs to understand the full instruction context before it reads the data. The example prompt has data set up as variables in the curly braces {} because it is part of a larger system calling from selected databases. Most people using a chatbot can do something like this instead:
Conclusion
When creating prompts, the key is simplicity with explicit directions. Vagueness allows AI to fill in any gaps and produce unpredictable output, i.e., internet slop. Precise writing is better than more writing. It is important to stress again: Prompting won’t eliminate hallucinations, it helps reduce them. See my related article on what comes after the prompt.