Prompt

A prompt is the input sent to an LLM to elicit a desired behavior - typically a combination of instructions, context, and examples.

Details

Modern LLM APIs structure prompts as a sequence of messages, each tagged with a role:

• System: sets the model's overall persona, behavioral constraints, and ground rules for the conversation. Placed at the beginning of the message sequence and typically treated as the highest-priority instructions by the model.
• Developer: carries instructions from the application developer (for example, tool definitions, output format requirements, or policy rules). Some API providers use developer messages as a distinct role separate from system; others treat system and developer as equivalent.
• User: contains the end-user's input--questions, requests, or data the model is asked to process.
• Assistant: represents the model's own previous responses. In multi-turn conversations, assistant messages are included in the prompt so the model can maintain coherence across turns.

These role distinctions are not purely labeling conventions; they are reinforced during post-training. Chat-oriented models are trained on data where system/developer instructions carry higher authority than user messages, which teaches the model to prioritize system-level constraints even when user messages conflict with them. This training-based priority is probabilistic rather than absolute--prompt injection attacks can still override it, especially when adversarial text is embedded in user-supplied content.

Because models treat prompts as their primary source of behavioral guidance, prompts function as a product's "behavior layer"--changing a prompt can shift outputs as much as changing the LLM itself. Production prompts are typically authored as prompt templates--parameterized structures with placeholders for dynamic content--and rely heavily on instruction following for consistent formatting, tone, and policy adherence.

Prompt management treats these templates as versioned artifacts: validated with evals (including regression tests from production failure cases) and monitored with observability that captures template versions, key inputs, outputs, latency, cost, and safety signals for debugging and drift detection.

Prompt design is one part of context engineering, which manages everything that goes into the LLM's context window and how it is structured under a given context size. The iterative practice of refining prompt wording, examples, and format is known as prompt engineering.