Prompt Structure

Prompt structure

A prompt is read in order, and models weight the beginning and end more than the middle. Structure decides what the model treats as instruction, what it treats as data, and what it thinks the output should look like, which is most of prompt quality.

Method

1. Lead with the task in one sentence. What to produce, for whom. A model that reads three paragraphs of context before learning its job interprets the context without a frame.
2. Separate instructions from material explicitly. Delimiters or tags around provided content, with a statement that it is data rather than instruction, which also blunts injection (see llm-guardrails).
3. Put the output specification near the end. The requirement closest to generation has the strongest effect on format (see output-format-control).
4. Order constraints by importance. Models drop instructions under load, and the ones buried mid-list go first.
5. Keep one prompt to one task. A prompt asking for a summary, a translation, and a critique does all three worse than three prompts.
6. Use consistent internal vocabulary. Referring to the same thing three ways within a prompt creates ambiguity the model resolves arbitrarily.
7. Cut everything that does not change the output. Politeness, preamble, and restated context dilute the instructions that matter.

Boundaries

Structure improves reliability; it cannot make a model capable of a task beyond it. Optimal ordering differs somewhat between model families, so a prompt tuned to one may need adjustment (see prompt-testing). Very long prompts degrade regardless of structure (see context-placement).