Prompt Template Patterns

# Prompt Template Patterns

Reusable patterns for structuring prompts. Covers delimiter conventions, system prompt architecture, few-shot formatting, output schema design, and provider-specific adaptations.

## Delimiter Conventions

Delimiters separate instructions from data from examples. Critical for both clarity and injection defense.

### Recommended Delimiters by Purpose

| Purpose | Delimiter | Example |
|---|---|---|
| **Sections** (role, context, instructions) | XML tags | `<system>`, `<context>`, `<instructions>` |
| **Data blocks** (user input, documents) | XML tags or triple backticks | `<user_input>...</user_input>`, ` ```...``` ` |
| **Examples** (few-shot) | Numbered with headers | `Example 1:`, `Input:`, `Output:` |
| **Output format** | Code block with schema | ` ```json { ... } ``` ` |
| **Internal reasoning** | XML tags | `<thinking>...</thinking>` |
| **Separators** (between sections) | `---` or blank lines | Visual separation for readability |

### XML Tags — Best Practice

XML tags are the most reliable delimiter for LLMs. They:
- Are unambiguous — clear open/close boundaries
- Nest naturally — `<context><document>...</document></context>`
- Are injection-resistant — attacker must match exact tag names
- Are model-agnostic — all major providers handle XML tags well