Clear instructions

Concept

Most poor AI output is not a model failure. It is an instruction failure. The model is doing its best with a prompt that did not actually specify the answer you wanted.

A clear instruction does four things at once. It states the task in one sentence. It names the audience or context. It specifies the output format (length, structure, tone). And it lists what to avoid.

Consider the difference between "write something about email marketing" and "write three subject lines for a sales onboarding email aimed at small-business owners; keep each under fifty characters and avoid the word free." The second prompt removes most of the room for misinterpretation. The model still has creative latitude inside those constraints, but the constraints are doing the heavy lifting.

A useful test: read your prompt back as if you were a new contractor seeing the task for the first time. Could you complete it without asking questions? If not, you are leaving the model to guess, and it will guess toward the safest, blandest answer in its training data.

Three traps to avoid. First, stacking too many instructions in one prompt without ordering them, which causes the model to weight them equally when you meant some to dominate. Second, using soft words like "good" or "engaging" without defining what good means in your context. Third, asking for output and then immediately constraining it ("write a long article but keep it short"), which forces the model to pick a side and rarely picks the one you wanted.

The rest of this resource builds on this foundation. Role prompting, structured output, few-shot examples. All of them are specialised forms of being clearer about what you want.