Prompt

A prompt is input supplied to guide a model's output for a task; it may include an instruction, question, example, context, or constraint [1].

A prompt tells an AI system what the user or product wants. It can be a short question, a long task brief, a set of examples, a pasted document, a format rule, or a combination of these. In many products, the visible user message is only part of the complete prompt-like context seen by the model.

A product may combine user input with system-level instructions, examples, retrieved passages, conversation history, tool results, and output constraints. Good prompt design can make output clearer and more useful. It does not guarantee truth, safety, source support, or freedom from hallucination.

Prompting is one of the easiest ways for ordinary users to improve AI output. A vague request such as "write this better" gives little direction. A stronger prompt names the task, audience, source material, style constraints, and success criteria. That helps the model produce an answer closer to the user need.

Prompting also matters because many AI products are built around hidden prompt assembly. The application may add policy instructions, retrieved knowledge-base passages, examples, and formatting rules before calling the model. A user should not assume the visible chat message is the entire request.

The prompt-learning survey "Pre-train, Prompt, and Predict" describes prompting as a broad NLP pattern where task information is expressed through prompts rather than only through traditional task-specific training setups [1]. For Richbay users, the practical lesson is smaller: prompts guide generation during use; they are not the same as training and do not update the model parameters by themselves.

A practical prompt usually has several parts. The task says what to do. The context gives material to use. The audience says who the answer is for. Constraints set length, format, tone, or exclusions. Examples show the desired pattern. A review instruction asks the model to flag uncertainty or separate claims from suggestions.

In a product, prompt design can happen at several layers. A system instruction may set role and safety boundaries. Application-level instructions may define workflow. Retrieved passages may provide evidence. Tool results may add current data. The user message may ask the visible question. The model output is shaped by the assembled request, not only by the last sentence the user typed.

Prompt design is separate from training. Training changes model parameters before use. Prompting gives task-specific context during a request. A prompt can make a model behave differently in that interaction, but it should not be described as teaching the underlying model in a permanent human-like way.

Imagine a user wants an AI assistant to summarize a workplace policy. A weak prompt says, "Summarize this." A better prompt says, "Summarize the supplied policy for frontline support staff in five bullets. Use only the policy text. Flag any rule that depends on country, customer plan, or manager approval."

The stronger prompt gives task, audience, source boundary, format, and risk flags. The output may be more usable, but the user still needs to check the policy. If the assistant says a refund is allowed, the source must actually support that claim. Prompt quality improves direction; source verification checks truth.

The biggest misconception is that prompt engineering can force truth. A clear prompt can reduce ambiguity, ask for citations, request uncertainty, and discourage unsupported claims. It cannot guarantee that the model will know, retrieve, or use the right evidence.

Another misconception is that prompts are only user-written chat messages. In real products, prompts may be assembled by software. The user may see one box, while the model receives system instructions, retrieved context, examples, previous turns, and tool outputs.

Prompting can create privacy and security risks. Users may paste sensitive data into a tool without checking policy. Attackers may hide instructions in documents, emails, webpages, or retrieved passages. OWASP prompt injection guidance describes indirect attacks where malicious instructions appear in external content processed by an AI system [2].

NIST also treats generative AI outputs, misuse, and reliability as risks that need governance and evaluation, not only better wording [3]. A prompt can ask for careful behavior, but a product still needs source checks, permissions, monitoring, and human review for high-stakes tasks.

Before sending a prompt, ask: What task do I want? What source material should the model use? What should it avoid? What format do I need? Is sensitive data included? Does this task need a primary source, a human expert, or a second check?

After receiving the answer, ask: Which claims matter? Which claims are supported by supplied sources? Did the model follow the constraints? Did it invent missing details? Would the answer cause harm if wrong? Good prompting starts the work. Verification decides whether the work is trustworthy.

Continue with Context Window for request capacity, AI Assistant for product-level use, and Source Verification for checking the resulting claims.