XML Tags Don't Help Short Prompts — Here's When They Actually Matter (2026)

The Conventional Wisdom Every prompt engineering guide says the same thing: wrap your prompt sections in XML tags. , , . Anthropic recommends it. OpenAI recommends it (with markdown headers). The internet treats it as best practice. But best practices need boundaries. When does this actually matter? And what happens when you apply structural overhead to prompts that don't need it? I ran the experiment. The answer is instructive. Setup Model: Claude Sonnet 4.5 Task: Extract 7 structured fields fr