Background Information & System Prompts

What Goes in Background Information?

Background information is static context that doesn't change between requests. This includes:

• Document structures and schemas
• Domain rules and definitions
• Reference data and lookup tables
• Process descriptions
• Constraints and limitations

Why Use System Prompts?

System prompts are ideal for background information because:

• Persistence - Stays constant across the conversation
• Caching - Can be cached to reduce latency and costs
• Separation - Keeps dynamic content clean
• Efficiency - Model learns the context once

The Car Accident Form Example

In Anthropic's demo, they provided complete form structure:

<form_structure>
  The Swedish car accident form ("Skadeanmälan") has:

  - Title: "Skadeanmälan vid trafikolycka"
  - Two columns: Vehicle A (left) and Vehicle B (right)
  - 17 numbered rows with checkboxes

  <row_definitions>
    <row number="1">Was parked/standing still</row>
    <row number="2">Was starting to move/leaving parking</row>
    <row number="3">Was parking</row>
    <row number="4">Was exiting a parking lot/private road</row>
    <row number="5">Was entering a parking lot/private road</row>
    <row number="6">Was entering a roundabout</row>
    <row number="7">Was driving in a roundabout</row>
    <row number="8">Was striking the rear of another vehicle</row>
    <row number="9">Was driving in the same direction in a different lane</row>
    <row number="10">Was changing lanes</row>
    <row number="11">Was overtaking</row>
    <row number="12">Was turning right</row>
    <row number="13">Was turning left</row>
    <row number="14">Was reversing</row>
    <row number="15">Was driving on the wrong side of the road</row>
    <row number="16">Was coming from the right (at intersection)</row>
    <row number="17">Did not observe traffic sign/signal</row>
  </row_definitions>

  <form_conventions>
    - Checkboxes may be marked with X, checkmark, or circle
    - Markings may be imprecise (human handwriting)
    - Both vehicles should have at least one box checked
    - The sketch section shows a hand-drawn diagram
  </form_conventions>
</form_structure>

What This Achieves

Prompt Caching Benefits

When using the API with prompt caching:

Static Background (Cached)
├── Form structure definition
├── Row meanings
├── Interpretation guidelines
└── ~2000 tokens (cached, fast)

Dynamic Content (Per Request)
├── Actual form image
├── Specific task instructions
└── ~500 tokens (new each time)

• Reduced latency (cached portion instant)
• Lower costs (cached tokens discounted)
• Consistent behavior (same background always)

Structuring Background Information

Use XML Tags for Organization

<background>
  <domain>
    Insurance claims processing for car accidents
  </domain>

  <document_types>
    <type name="accident_form">
      Standardized form with checkboxes...
    </type>
    <type name="sketch">
      Hand-drawn diagram showing...
    </type>
  </document_types>

  <terminology>
    <term name="Vehicle A">The reporting party's vehicle</term>
    <term name="Vehicle B">The other party's vehicle</term>
    <term name="fault">Legal responsibility for the accident</term>
  </terminology>

  <rules>
    <rule>Both vehicles must have actions checked</rule>
    <rule>Fault is determined by traffic laws</rule>
    <rule>Rear-ender is typically at fault</rule>
  </rules>
</background>

Progressive Detail

Start general, then get specific:

1. Domain overview (what business/field)
2. Document description (what you'll see)
3. Field definitions (what things mean)
4. Processing rules (how to interpret)
5. Edge cases (special situations)

What NOT to Put in Background

Avoid including:

• Information that changes per request
• The actual content to analyze
• Specific task instructions (those go in user prompt)
• Output examples (those are few-shot, separate section)

Real-World System Prompt Structure

You are an AI assistant for [COMPANY] that [ROLE DESCRIPTION].

<background_knowledge>
  [Static domain information]
  [Document structures]
  [Terminology definitions]
  [Business rules]
</background_knowledge>

<behavior_guidelines>
  [How to handle uncertainty]
  [Quality standards]
  [Ethical constraints]
</behavior_guidelines>

When given [INPUT TYPE], you will [TASK DESCRIPTION].

Example: E-Commerce Product Analyzer

<system_prompt>
You are a product data analyst for an e-commerce platform.

<product_categories>
  <category id="electronics">
    Includes: phones, laptops, tablets, accessories
    Key attributes: brand, specs, warranty, compatibility
  </category>
  <category id="clothing">
    Includes: shirts, pants, shoes, accessories
    Key attributes: size, color, material, care instructions
  </category>
</product_categories>

<quality_rules>
  - Titles must be under 200 characters
  - Descriptions must mention key features
  - Prices must be in USD
  - Images must show actual product
</quality_rules>

<common_issues>
  - Keyword stuffing in titles
  - Missing size information
  - Stock photos instead of actual product
  - Inconsistent pricing
</common_issues>

Analyze product listings for quality issues and compliance.
</system_prompt>

Key Takeaways

• Background info is static - Doesn't change between requests
• System prompts are ideal - Persistent and cacheable
• Structure with XML - Clear organization helps the model
• Be comprehensive - Include everything the model needs to know
• Cache when possible - Reduces cost and latency

Next Lesson

Next, we'll explore Structuring with XML Tags - how to organize information within prompts for maximum clarity.