Structuring with XML Tags

Why Claude Loves Structure

Claude (and other LLMs) perform better with structured input because:

• Clear boundaries - Knows where information starts and ends
• Semantic labels - Understands what information represents
• Easy reference - Can refer back to specific sections
• Consistent parsing - Reliable pattern matching

XML Tags vs. Other Formats

Basic XML Tag Usage

Wrapping Content

<user_input>
The text that the user provided goes here.
It can span multiple lines.
</user_input>

Labeling Sections

<task>Summarize the following article</task>

<article>
Article content goes here...
</article>

<requirements>
- Keep summary under 100 words
- Include main thesis
- List key points
</requirements>

Nesting for Hierarchy

<context>
  <domain>Healthcare</domain>
  <audience>Medical professionals</audience>
  <purpose>Treatment recommendation</purpose>
</context>

Naming Conventions

Use clear, descriptive tag names:

<!-- Good -->
<customer_feedback>...</customer_feedback>
<product_description>...</product_description>
<analysis_output>...</analysis_output>

<!-- Bad -->
<data>...</data>
<text>...</text>
<output>...</output>

• Use snake_case for multi-word tags
• Be specific about content
• Avoid generic names
• Match output tags to input tags

Real-World Examples

Example 1: Document Analysis

You are a document analyst. Analyze the following document.

<document type="contract">
SERVICE AGREEMENT

This agreement is entered into between Company A ("Provider")
and Company B ("Client") on January 15, 2025...
</document>

<analysis_requirements>
  <extract>
    - Parties involved
    - Effective date
    - Key obligations
    - Termination clauses
  </extract>
  <format>
    Provide a structured summary with sections for each item
  </format>
</analysis_requirements>

Output your analysis in <analysis> tags.

Example 2: Code Review

Review the following code for issues:

<code language="python">
def calculate_total(items):
    total = 0
    for item in items:
        total = total + item.price
    return total
</code>

<review_criteria>
  - Performance issues
  - Code style
  - Potential bugs
  - Best practices
</review_criteria>

Provide feedback in <code_review> tags.

Example 3: Multi-Input Analysis

Compare these two product descriptions:

<product_a>
  <name>Widget Pro X</name>
  <price>$299</price>
  <description>
    Professional-grade widget with advanced features...
  </description>
</product_a>

<product_b>
  <name>Widget Basic</name>
  <price>$99</price>
  <description>
    Entry-level widget for beginners...
  </description>
</product_b>

<comparison_criteria>
  - Value for money
  - Feature set
  - Target audience
  - Marketing clarity
</comparison_criteria>

Using Tags for Output

Requesting Tagged Output

Analyze the sentiment of this review:

<review>
I loved this product! The quality exceeded my expectations,
though shipping was a bit slow.
</review>

Respond with:
<sentiment_analysis>
  <overall>positive/negative/mixed</overall>
  <score>1-10</score>
  <positive_aspects>...</positive_aspects>
  <negative_aspects>...</negative_aspects>
  <summary>...</summary>
</sentiment_analysis>

Benefits of Tagged Output

• Easy parsing - Extract specific fields programmatically
• Consistency - Same structure every time
• Validation - Check for required elements
• Integration - Feed into downstream systems

Parsing XML Output

In your application:

// Example: Extracting tagged content
function extractTag(response, tagName) {
  const regex = new RegExp(`<${tagName}>([\\s\\S]*?)</${tagName}>`);
  const match = response.match(regex);
  return match ? match[1].trim() : null;
}

// Usage
const sentiment = extractTag(response, 'overall');
const score = extractTag(response, 'score');

Common Patterns

Input-Output Mirroring

<!-- Input -->
<user_query>How do I reset my password?</user_query>

<!-- Request output in matching structure -->
Please respond with:
<response>
  <answer>...</answer>
  <steps>...</steps>
  <related_topics>...</related_topics>
</response>

Conditional Sections

<instructions>
If the query is a question, include <answer> tags.
If the query requires steps, include <steps> tags.
If uncertain, include <clarification_needed> tags.
</instructions>

Metadata Tags

<output>
  <metadata>
    <confidence>high/medium/low</confidence>
    <sources_used>...</sources_used>
    <processing_notes>...</processing_notes>
  </metadata>
  <content>
    The actual response content...
  </content>
</output>

Key Takeaways

• XML tags add structure - Claude understands and respects boundaries
• Use descriptive names - Clear labels help the model understand context
• Be consistent - Use the same tags throughout your prompt
• Request tagged output - Makes parsing reliable
• Nest for hierarchy - Group related information

Next Lesson

Next, we'll explore Few-Shot Learning with Examples - how to demonstrate expected behavior through input/output pairs.