Prompt

A prompt is whatever text you feed an LLM to get a response. In ChatGPT or Claude.ai, the prompt is the message you type plus, behind the scenes, the system prompt and prior chat history. In an API call, you control the whole thing — system message, user message, assistant turns, tool definitions — and the model sees all of it as one big input. "Prompt engineering" emerged in 2022-2023 as a real craft when people realized that small wording changes (asking the model to "think step by step", giving it a persona, providing examples) could dramatically change output quality. Some of that craft is now obsolete — frontier reasoning models in 2026 don't need "think step by step" prompted, they do it natively. But the core skill — writing clear, specific, well-structured instructions — is more valuable than ever, especially for agents where a bad prompt cascades through dozens of tool calls. The line between "prompt" and "context" is fuzzy. When you paste a 50-page PDF into Claude and ask a question, is the PDF part of the prompt? Practically, yes — everything you put in the model's input window is the prompt. Some people distinguish "instructions" (what you want done) from "context" (data the model needs to do it), but the model just sees one long sequence of tokens.