What it means
A prompt is whatever text you feed an LLM to elicit a response: a question, an instruction, a few examples, a piece of context for the model to work with. The model reads the prompt and produces a continuation. Everything an LLM does is downstream of the prompt.
Prompts can be a single sentence ('Summarise this email') or many pages ('Here is our brand voice guide; here are 50 examples of correct responses; here is a customer message; reply in our voice'). The longer and more detailed the prompt, generally the more controllable the output.
Why it matters
Prompt quality is the difference between an AI agent that sounds generic and one that sounds like your brand. Two systems running on the same LLM can produce wildly different results based purely on prompt design.
Prompt engineering is also where most production AI work actually happens. Switching LLM models is rare; iterating on prompts to fix specific failure cases is daily.
Example
A property agency's AI agent initially has a generic prompt ('Help customers with property questions'). Replies are accurate but bland. The team rewrites the prompt with the brand voice, common-question examples, and explicit guardrails ('Never quote prices without checking the unit-sheet tool'). Same LLM, vastly better replies.