In one sentence

A prompt is the instructions and information you give a model to produce a response.

The simple model

A prompt is what the model sees when it does its work: the task, any rules you want it to follow, and the information it should use. There is often a standing part, sometimes called a system instruction, that sets the model's role and rules, and a specific part for each request.

Writing a clear prompt is a real skill, and it matters. But it is important to understand what a prompt is and is not.

A prompt guides; it does not guarantee

The model treats a prompt as strong guidance, not as an unbreakable rule. Because output is probabilistic, a model may not always follow every instruction, especially under unusual or crafted inputs.

This has a direct consequence for design: anything that must always hold cannot rely on the prompt alone. "Never do X" written in a prompt is a request the model usually honours. If X must never happen, enforce it with a guardrail in the surrounding system.

What belongs in a prompt

  • The task and the desired form of the answer.
  • The context and sources the model should use, kept relevant and within the context window.
  • Instructions to say when it does not know, rather than inventing an answer.
  • Tone and format guidance.

What does not belong in a prompt as your only line of defence: hard limits, prohibited actions, and anything whose failure would be costly. Those need enforcement, not instruction.

A small example

A team wrote "never share internal pricing" in the prompt and considered the risk handled. Under an unusual question, the model shared it anyway. The reliable fix was not a stronger sentence; it was keeping internal pricing out of what the model could access in the first place. The prompt guided; the access control enforced.

What this does not solve

A good prompt improves results; it does not make them deterministic, correct or safe by itself. Prompts are one input to a probabilistic system. Guardrails, validation and review are what turn a helpful model into a dependable workflow.