Prompt

Google’s glossary defines a prompt broadly: any text given to a large language model to steer how it behaves, from a short phrase to an entire novel (Google). Image, audio and video models take prompts too, and many models now accept images or files as part of one.

What you type is often only part of it. A chat app usually adds a system prompt with standing instructions, the earlier turns of the conversation, retrieved documents and descriptions of available tools. The model sees all of it as one input, measured in tokens and limited by the context window.

A concrete comparison: “Summarize this” with meeting notes pasted below gets a generic summary. “Summarize these notes for a VP who missed the meeting: three bullets on decisions, then owners and deadlines, under 100 words” usually gets a more useful one, because it says who the output is for, what to include and what shape it should take. Adding worked examples is few-shot prompting; giving none is zero-shot prompting. Both rely on in-context learning.

Prompts matter because they are the cheapest thing you control: no training, instant iteration. They are also fragile, since small wording changes can change results, so treat a prompt used in a product like code: version it, and test changes against a fixed set of examples. Anthropic’s guidance starts from success criteria and notes that some problems, such as latency or cost, are better solved by choosing a different model than by rewording (Anthropic). The Prompt Report catalogs 58 text prompting techniques for those who want to go deeper (Schulhoff et al.). Prompt engineering is the practice of designing prompts; untrusted text inside a prompt can also be an attack, covered under prompt injection.