Input Format

Input format is about how you physically arrange a prompt: the order of its sections, the delimiters that separate instruction from data, and the markup you use to label each part. The same words, restructured, can produce measurably different results — research on prompt formatting has shown double-digit accuracy swings between plain text, Markdown, and structured layouts of identical content, and the best format even varies by model family.

For developers this is a free performance lever and a safety mechanism at once. Clear structure helps the model distinguish what you’re telling it to do from what it should operate on — and that boundary is exactly what prompt injection attacks try to blur. A user-supplied document dropped inline with your instructions is an invitation for embedded text like “ignore previous instructions” to be obeyed; the same document fenced inside labeled delimiters is far easier for the model to treat as inert data.

In practice, you’ll adopt a consistent house format: instructions first, then labeled data blocks using XML-style tags (<document>...</document>, <user_query>...</user_query>) or triple-backtick fences, with Markdown headers for multi-section prompts. Put the most important instruction at the top and restate critical constraints near the end of long prompts. When you switch model providers, re-test your formatting — structure that’s optimal for one model is merely good for another, and your eval suite will tell you the difference.