xml-tag-structure

XML Tag Structure

Overview

This skill provides guidelines and tools for utilizing XML-like tags to structure prompts, format LLM output payloads, and robustly parse structured text data. It helps maintain clear delimiters for instruction blocks and data boundaries, even when dealing with malformed or truncated text.

File Tree

xml-tag-structure/
├── SKILL.md                     # This file. Skill definition and workflow guide.
├── references/
│   └── xml-guidelines.md        # Comprehensive best-practices reference for XML tags.
└── scripts/
    └── parse_xml.py             # CLI tool for extracting tag contents from text.

Workflow Decision Tree

To handle XML tags in LLM workflows, follow this process:

1. Structuring Prompts: Wrap instructions, context, input variables, and output schemas in clean, snake_case XML tags.
2. Structuring Outputs: Define specific tags for the model to output, using CDATA sections for nested code or raw markup blocks.
3. Parsing Output Payloads: Run robust parsing logic to extract tags, handling potentially truncated responses and malformed markup.

---

1. Structuring Prompts

When designing system prompts, wrap different types of information in explicit tags. This helps the LLM distinguish instruction from context or user input, preventing prompt injection and instructions leakage.

Standard Tag Categories

• <system_instructions>: Define the agent's identity, rules, and core directives.
• <context>: Provide background information, schemas, file contexts, or system state.
• <examples>: Enclose few-shot examples (each inside <example>).
• <user_input>: Enclose the current input or request to process.

Rules for Prompt Tags

• Use snake_case or kebab-case for tag names (e.g., <file_content> or <file-content>).
• Keep tag names descriptive and singular/plural consistent (e.g., <item> inside <items>).
• Avoid mixing tags with standard markup or styling unless explicitly wrapping inline values.