Advanced System Prompt Design: XML Tagging, Delimiters & Reasoning Scaffolds

How to structure system instructions that eliminate drift, prevent prompt injection, and yield deterministic output schemas.

Prerequisites

Basic familiarity with LLM chat interfaces

Step 1: Step 1: Enclose Untrusted User Input in Semantic Tags

Models can be steered by malicious user prompts if they blend with instructions. Always wrap context and user input inside explicit XML boundaries.

Step 2: Step 2: Specify Thinking Protocols and Output Schemas

Instruct the model to emit its internal analysis within <scratchpad> tags before delivering the final response inside <response> JSON tags.