A prompt uses XML-style tags around its instructions, Markdown headings around its context, and blank lines to separate its examples. It works, but unpredictably. What does the guidance say about this?

The principle — A delimiter works because it is recognisable as a boundary. Three conventions in one prompt mean the model is inferring which marks are structural and which are content.

Why the key is correct — Google's guidance is to use consistent structure, employing clear delimiters to separate the parts of a prompt, and states directly that you should choose one format and use it consistently within a single prompt. Anthropic makes the same point for tag naming, advising consistent and descriptive tag names across prompts. Distinguishing sections is the job of the names inside the convention, not of switching conventions.

Why the others are wrong — Mixing formats adds a second convention to decode rather than aiding separation. Structure is not merely presentational; Google Cloud says it helps the model parse information that is otherwise complete. And the need for boundaries has nothing to do with whether the pasted material happens to be code.

Remember this — One convention per prompt. Vary the tag names, never the tagging style.

Sources — Google, Prompt design strategies.