Prompt Design for Language Models

Most people first meet a language model through a chat window. You type a question, and text comes back that looks like an answer. It is natural to assume the model is listening to you the way a person would, understanding your request and replying to it. That assumption shapes how people use these systems, and it is also the source of many surprising results. This article starts from that assumption, tests it, and then builds a more accurate picture of what the model is doing when it writes text.

The central idea is that a language model can be treated as a completion machine. It receives a piece of text, called the prompt, and produces the text that most plausibly follows. Everything else in this article is a consequence of that one framing.

Before reading on, make a prediction. Imagine a model that does nothing except continue whatever text it is given. You give it a question and nothing else. What does it do with that question?

The surprise in that prediction is the point. A model that writes fluent replies in a chat window can still, under the same mechanism, produce a question in return. The fluency was never the same thing as understanding that you asked for an answer.

Think of the prompt as the start of a passage that someone has left unfinished. The model's job is to supply the continuation. It does this by estimating which pieces of text are likely to come next, given everything in the prompt. Those estimates are probabilities, so the model does not produce one fixed answer by rule. It produces a continuation that is likely given the input.

The same behavior works with a prompt that is not a famous quotation. If you begin a sentence such as 'The lighthouse keeper climbed the stairs and', the model's continuation depends on which words most commonly follow sentences like that. The model does not need to know a story about lighthouses to extend it. It needs only to predict the next text that fits.

The branches in the diagram are not all equally likely, and the source does not give a single rule that picks one of them. The model's output is one draw from a set of plausible continuations. Run the same prompt twice and a different branch can appear, because the output is a probabilistic prediction rather than a fixed lookup.

A completion does not stop being part of the input once it is written. Each new piece of text the model produces becomes part of the text it reads next. The model then predicts the continuation of that longer passage, and so on. The prompt is the starting point of a loop, not a single question-and-answer step.

This loop explains why a small change in the prompt can change the whole output. The first continuation the model writes sets the context for everything after it. A prompt that leads toward a list will tend to produce a list, and a prompt that leads toward a question will tend to produce a question.

The misconception to correct is this: a language model that produces fluent replies must be engaging with your request as a conversation partner. Completion is a different activity. A conversation requires the model to recognize that you asked something and to respond to that request. A completion does not require that at all.

The source notes that the step of shaping a model to respond appropriately to a user's request is handled in a later stage called post-training. The basic completion behavior comes first, and conversational behavior is a layer that has to be taught on top of it. When you see a model answer your question, you are seeing that layer working on top of the completion mechanism.

A completion is a prediction, and predictions can be wrong. A continuation can read smoothly and still be factually incorrect, because fluency and correctness are separate properties. The model is choosing text that fits the pattern of what usually follows, and a pattern can look right while being false.

The source describes this probabilistic character as what makes language models both exciting and frustrating to work with. The same mechanism that makes them flexible also makes their output uncertain. A model can produce a translation, a summary, or a label that looks authoritative, and still be a plausible guess rather than a verified result.

Once you accept that a model completes text, many jobs can be restated as completion problems. The source highlights translation, summarization, coding, and solving math problems, along with spam classification. The trick is to write the prompt so that the most likely continuation is the output you want. The task is turned into a text prefix whose natural ending is the answer.

The translation row follows the source directly. The prompt 'How are you in French is' leaves a gap that a plausible continuation can fill with 'Comment ça va'. The model is not given a translation rule to execute. It is continuing a pattern in which a phrase is followed by its French rendering, so the translation appears as the completion.

Spam detection follows the same pattern. A prompt that asks 'Is this email likely spam?' and then includes the email text sets up a passage where 'Likely spam' is the plausible continuation. Seen this way, the language model becomes a spam classifier, because the label it writes is simply the continuation that fits the question.