A user's request to an AI application such as a chatbot. Commonly called a "prompt" or "text prompt," the way the prompt is formulated determines the outcome that is generated. Even small changes in the prompt can make a huge difference in the results.
From Simple to Gigantic
AI prompts can be a very short sentence; however, they can include uploaded files of data, which means a prompt can be "thousands" of words. Prompts are often many pages long because they describe the format of the inputs to the job and the requested outputs. Even ancient file formats can be described to a chatbot in order to convert them to a modern format as well as customize the output for a particular job. See
text prompt,
AI prompt engineer and
generative AI.
The Multi-Turn Tax
Because uploaded files are part of the prompt, whenever the user asks multiple questions that pertain to those same files, the chatbot re-reads the uploaded files again and generates an answer starting all over again. If users pay the chatbot company based on number of output tokens, each answer may cost as much as the first, or the user may run into token limits, depending on the subscription plan they have.
System Prompts Are Different
Also used at inference and not training, a system prompt modifies the behavior of a model. For example, it might request avoiding certain types of content. However, system prompts are generally entered by the AI developer, although some models allow user system prompts to obtain results in a prescribed format; for example, "explain as simply as possible." In either case, system prompts are generally temporary modifications to the AI model and differ from the user requests that generate answers. See
system prompt leakage,
chain of thought prompting,
neural network and
AI model.