A prompt is the input you give an AI model: the question, instructions, context and examples that the model conditions its response on. Since an LLM works by predicting what plausibly comes next, the prompt is quite literally the start of the text it’s continuing — change the prompt, and you change what “next” looks like. It’s the primary steering mechanism you have.
The layers of a prompt
What the model actually sees is usually assembled from parts:
- System prompt — standing instructions from the developer that frame the whole session: role, rules, tone, tool definitions. Users don’t see it, but it wins conflicts with user requests by design. (Vendors’ production system prompts run thousands of words — we analyze leaked ones on our model pages.)
- User prompt — what you type: the task itself.
- Context — documents, retrieved snippets, prior conversation turns and tool outputs riding along in the context window.
What prompt engineering actually is
Prompt engineering is the craft of phrasing and structuring all of that so the model reliably does what you want. The techniques that consistently pay off are unglamorous: be specific about the task and output format, show a worked example or two (“few-shot”), give the model room to reason before answering, structure long inputs with clear delimiters, and say what to do rather than only what to avoid. The skill matters less per-request than it used to — frontier models infer intent better — but it compounds in production, where one prompt template runs thousands of times and every wasted token is billed.
Where to go deeper
See real production system prompts dissected on model pages like GPT-5.5 and Claude Opus 4.8, or check what your prompts cost with the Token Calculator.