Prompt

Concept

Also known as: Input, User Prompt

Text input that forms the starting point for a language model’s computation

A prompt is the text input with which a user starts or continues a language model’s computation. This input becomes part of the current context on which the model calculates its answer. Before the model generates a single token of its own, the entire basis for every subsequent computation has already been established. A prompt is therefore not a command in the human sense, but the starting point of a token-by-token computation.

What is a Prompt?

A Prompt is the text input with which a user starts or continues a language model’s computation. This input becomes part of the current context on which the model calculates its response. Before the model generates a single token of its own, the entire basis for every subsequent computation has already been established. A prompt is therefore not a command in the human sense, but the starting point of a token-by-token computation.

How it works

The prompt is first tokenized and becomes the beginning of the context window. Every token generated by the model is calculated based on all previous tokens, including the entire prompt. The model responds to the actual sequence of tokens present in the context, not to the human intention behind them.

Why different wording produces different answers

Because the model has no understanding of meaning independent of the specific context, different wording with the same intended meaning often changes the answer. Different phrasings become different token sequences that were associated during training with different typical continuations.

A prompt is therefore not a question addressed to an entity that understands meaning. It is the starting point of a statistical computation. A language model does not remember a prompt beyond the current conversation—every new conversation starts without reference to previous prompts unless they are explicitly provided again.

Why Prompts have practical implications

Wording: More precise and specific wording often leads to more suitable answers because it was associated with more appropriate continuations during training.
Context usage: A long prompt consumes part of the available context window before the model even begins generating a response.
No guarantee: A prompt does not force a particular behavior—it only shifts which continuations become more likely.

Typical example

“Explain Docker.” and “What is Docker?” mean essentially the same thing to a person, yet they often produce answers of different length or style because each wording was associated during training with different typical continuations.

Sources

Hauptquellen

Vertiefung

Attention Is All You Need (Transformer Paper) [extern]

Sources archived on: 2026-08-02