Prompting
Prompt
Also known as: input, instruction
The text you give an AI model to tell it what you want; its input.
A prompt is simply what you type to an AI: your question, request or instruction. "Write a birthday message for my dad" is a prompt. The AI reads your prompt and responds to it. The clearer and more complete your prompt, the better the answer, which is a skill in itself.
A prompt is the input you provide to a language model: instructions, questions, context and any examples. The model's entire response is a continuation of that prompt, so what you include, and leave out, shapes the result.
Parts of a good prompt
- Instruction: what you want done.
- Context: the background facts or documents the model needs.
- Examples: optional worked examples that show the desired style or format.
- Constraints: length, tone, audience, output format.
Many chat systems also use a hidden system prompt set by the application, which defines the assistant's role and rules before your message is added. The deliberate craft of writing prompts is called prompt engineering.
A prompt is the conditioning context supplied to a model at inference; formally, generation samples from the distribution over continuations conditioned on the prompt tokens. In chat models the prompt is usually structured into roles, most commonly a system message (application-level instructions and persona), user messages, and prior assistant turns, all serialised into the context window.
Because a pretrained model performs in-context learning, the prompt can specify a task through instruction alone (zero-shot) or through demonstrations (few-shot), without weight updates (Brown et al. 2020). Consequently prompt content, ordering and framing are load-bearing: they determine task specification, grounding and output structure. This is why practitioners distinguish the prompt (a single input) from prompt engineering (the systematic practice of designing and evaluating prompts), and increasingly from context engineering, the broader problem of deciding what information to place in the window at all.