What is Prompt?
Also called input, query.
A prompt is the complete input given to a model for one request. It usually combines several parts: instructions describing the task, any supplied context or documents, examples, the conversation history, and the user's actual question. The model receives all of this as one sequence of tokens, with no inherent boundary between the parts beyond what its training taught it to expect.
Because a prompt is a single sequence, structure has to be signaled rather than enforced. Headings, delimiters, and role labels help a model separate instructions from data, but they are conventions learned during training, not hard boundaries in the system. This is precisely why prompt injection works: text inside a supplied document can read to the model as further instructions.
Prompts are also the main mechanism for supplying current information. A model's own knowledge stops at its training cutoff, so anything recent, private, or specific to one user has to be placed into the prompt at request time, typically by retrieving the relevant passages from a store and inserting them before the question the user actually asked.
Length carries costs in both directions. Everything in the prompt counts against the context window and is metered, so long prompts are slower and more expensive per request. Overly long prompts also dilute attention, and instructions buried in the middle of a large block are followed less reliably than the same instructions placed at the start or the end.
The word carries two meanings in practice, which causes confusion in discussions. Casually, prompt means whatever the user typed into a box. Technically, it means the entire assembled input including the system prompt, retrieved context, tool definitions, and prior turns. In most real applications the typed portion is a small fraction of what the model actually receives.
Key points
- The full assembled input, not just what the user typed.
- Instructions and data share one sequence; boundaries are convention only.
- The only route for information created after the training cutoff.
- Every prompt token counts against the context window and is metered.
In practice
A user types 'Is this covered?' into a support tool. What reaches the model is far larger: a system prompt defining the assistant's role and rules, three retrieved policy passages, the last six turns of conversation, a list of available tools, and finally those three typed words. The user's question might be one percent of the total tokens sent.