What a prompt actually is

Almost every prompt that works has three parts. Recognising them is most of what people mean by prompt engineering.

The three parts

Context is what the model needs to know that it cannot infer: the audience, the codebase, the constraints, the source text. Instruction is the task itself. Output shape is what should come back - the length, the format, the sections.

Most prompts that disappoint are missing the third. "Summarise this" has context and instruction, but no shape, so the model picks one. It usually picks a competent, generic one aimed at nobody.

Order changes the answer

Put the input last. Models weight instructions near the end of a prompt more heavily, and a clearly marked block at the bottom - after a line like --- text --- - is much less likely to be read as an instruction. This is also the first line of defence against prompt-injection, where text inside your input tries to act like a command.

So the working order is: context, instruction, output shape, then input.

Vague against specific

Compare these.

Make this text better please.

Rewrite this product description for a Slovak audience. Keep it under 60 words, keep the two technical specs exactly as they are, and drop the marketing adjectives.

The second is not more polite or more detailed for its own sake. Every clause removes a decision the model would otherwise make for you: who it is for, how long, what must not change, what to cut. That is the whole difference.

The rule that saves the most time

Give the model a way to answer nothing. A prompt that asks for three problems will produce three problems whether or not they exist, because the model completes the pattern it was given. Add "if there are none, say so in one line" and the invented findings stop.

What to do next

Take a prompt you already use and write down which of the three parts it is missing. Then read the few-shot page in the glossary - showing one worked example is usually faster than describing the shape in words.