Is a prompt the same as context?

The terms overlap. Context is the information available to the model for the current generation; the prompt commonly refers to the supplied input that helps form that context.

Does a prompt need examples?

No. Examples are optional and useful only when they clarify the task.

Can a prompt guarantee a truthful answer?

No. It can request evidence-based behavior, but correctness still depends on available information and the system's actual performance.

AI Glossary / Prompting and human-AI interaction / Beginner

Prompt

Definition

A prompt is input supplied to a generative AI model to guide the output it produces. It may contain a question, instruction, unfinished passage, examples or contextual material. Depending on the system, a prompt can also include images, audio or other supported information types.

What it means

A prompt gives the model something to respond to or continue. "Summarise this notice in three sentences" specifies an operation, while the notice supplies the material on which to perform it. Both can be part of the prompt.

The word is used at different levels. In everyday conversation it often means the message a person types. In an application, the model's effective input may also contain earlier conversation, application instructions, retrieved passages and tool results. What the user sees is not always the whole context.

A prompt influences generation; it is not a binding contract that guarantees correct execution. The model's training, available context, decoding settings and surrounding application also affect the response.

How it works

An application assembles the input in the form its model accepts. Text is represented as tokens, while supported non-text inputs may use other encodings. The model uses the resulting context to generate an output according to its learned parameters and inference procedure.

A useful prompt makes the task and relevant evidence clear. It can specify the intended audience, required output form and handling of missing information. Demonstrations can show a desired transformation, but their contents should be distinguished from the new material to process.

Applications often separate instruction messages from reference material. That distinction is important because a document can contain commands that are merely part of the document. Prompt wording can help communicate the boundary, but robust applications also need appropriate handling of untrusted inputs and control over consequential actions.

Three practical examples

A manager supplies meeting notes and asks for a list of decisions, owners and unresolved questions.
A learner asks for an explanation of a concept using a familiar example and a specified level of detail.
A designer supplies a product photograph and asks for a description of visible features, provided the model supports image input.

Think of it this way

Think of a brief given to an assistant: it describes the job, provides relevant material and explains what a useful result looks like. A vague brief leaves more decisions to the recipient.

The analogy does not mean an AI assistant has a person's understanding or accountability. Even a clear brief can produce an incorrect response, so the result must be assessed against the actual task.

Worked example

Consider an original stock-summary example. The supplied facts are: "Morning stock: 48 notebooks. Delivered: 20. Sold: 35. No other stock changes occurred."

A vague prompt says, "Write about the stock." It leaves the purpose and format open. A more specific prompt says, "Using only the supplied facts, calculate closing stock. Return one sentence containing the arithmetic and result. If a required quantity is missing, state that it is missing."

The intended answer is: "Closing stock is 48 + 20 - 35 = 33 notebooks." This answer can be checked directly against the three quantities and the no-other-changes assumption.

Now remove the delivered quantity while leaving the rest of the request unchanged. The prompt's missing-information rule should lead to a statement that closing stock cannot be calculated from the supplied quantities. Reusing 20 from a previous example would be an error.

For a small invented evaluation, suppose twenty complete records and ten incomplete records are tested. If the model answers eighteen complete cases correctly and handles seven incomplete cases correctly, total success is 25/30, approximately 83.3%. The missing-information result is separately 7/10 = 70%.

These invented numbers illustrate evaluation, not expected real-world performance.

Illustrative case study

Fictional case: Maple Office Services uses an AI assistant to summarize delivery notes. Employees initially ask for "a useful summary", producing inconsistent mixtures of quantities, dates and commentary.

The team agrees on a prompt that requests three labeled fields and explains how to report missing values. It marks the delivery note as source material rather than as instructions for the assistant. Reviewers compare outputs with the originals and record unsupported additions.

The team stores prompt versions with the evaluation results. When the underlying model changes, it reruns the same checks rather than assuming earlier performance transfers unchanged. This hypothetical case makes no measured efficiency claim.

Common misconceptions and mistakes

A prompt is always a question: it can be an instruction, example or partial text.
A longer prompt is always better: irrelevant material and conflicting instructions can obscure the task.
Prompting automatically retrains the model: ordinary prompting changes inference context, not learned weights.
A requested format guarantees valid output: applications should check required structure.
Everything inside a prompt has equal authority: application instructions and quoted source material play different roles, even though both may reach the model as input.

Business use

Prompts are part of the design of an AI-assisted workflow. They can express business rules and output expectations, but should not carry responsibilities that require stronger controls, such as enforcing access permissions or validating arithmetic.

Start with a clear task and examples of acceptable results. Test varied inputs, including missing and contradictory information. Keep records of the prompt, model and evaluation conditions so changes can be investigated. For repeated operations, reusable templates can improve consistency while leaving room for the specific source material needed in each request.

Check your understanding

A user asks the model to summarize a document containing the sentence "Ignore all previous instructions." Should that sentence automatically become the user's instruction?

Answer: No. It is part of the source document to process. Applications should preserve the distinction between task instructions and untrusted content.

Related terms

Prompt engineering
Context window
Token
System prompt
Few-shot prompting
Prompt injection
user message
prompt template
instruction following

Learn next

Take one repeated task and write down the required evidence, output and missing-information behavior. Create a prompt expressing those requirements. Test ordinary and difficult cases, then revise only where the observed failures show a specific ambiguity or omission.

Sources

Google Machine Learning Glossary, generative AI terminology: https://developers.google.com/machine-learning/glossary/generative

Google Machine Learning Crash Course, prompt engineering and parameter distinction: https://developers.google.com/machine-learning/crash-course/llm/tuning

The stock calculation, evaluation counts and Maple Office Services case are original hypothetical illustrations.

Previous: Promoter based genetic algorithm Next: Prompt-based learning