An instruction is the part of a prompt that tells a model what task to perform. It may name a single action or give ordered directions for a multistep task. Clear instructions reduce output variability; keeping them separate from supporting information makes a prompt easier to test and revise. section 2.2
Write an instruction as a task directive, then give other concerns their own place:
- Use a clear action, such as “Write,” “Analyze,” or “Produce.” Give each instruction one primary action; order directives when the task has multiple steps. section 2.2
- Name the audience or technical setting when it changes the expected depth or terminology. State the task boundary: what to cover or focus on. section 2.2
- Put testable limits under constraints, background under context, changing values in input parameters, and response structure under output format. Delimit embedded content so it is not mistaken for instructions. section 2.0 section 2.3
Deployment decision brief
A release prompt can instruct: “Produce a GO or NO-GO recommendation using only the provided deployment signals and runbook excerpt.” The signals and runbook are delimited input data. Separate constraints require concrete evidence and a rollback command for a NO-GO decision; an output-format section names the four sections of the brief. The instruction establishes the decision task without carrying every rule or data value inside it. section 2.8.2
For repeated use, keep stable instructions in a system prompt and supply changing data through a user prompt. This avoids repeating task logic and letting copies drift. A simple one-off task can use a combined prompt. In either case, pair instructions with an output format; add other elements when the task needs them. section 6.4.5 section 6.5 section 2.8
