Prompt Structure Fundamentals

The Anatomy of a Great Prompt

When working with LLMs via the API, you typically want to send one comprehensive message and get an accurate response on the first try. This requires a well-organized prompt structure.

The 10-Point Prompt Framework

Anthropic recommends this structure for building production-quality prompts:

1. Task Context       - What is the model here to do?
2. Tone Context       - How should the model respond?
3. Background Data    - Static information it needs
4. Dynamic Content    - The actual input to process
5. Detailed Instructions - Step-by-step guidance
6. Examples           - Demonstrations of expected behavior
7. Output Format      - How to structure the response
8. Task Reminder      - Reinforce key requirements
9. Important Guidelines - Critical constraints
10. Prefill           - Start the response

Framework Visualization

┌─────────────────────────────────────────────────────────┐
│                     SYSTEM PROMPT                        │
├─────────────────────────────────────────────────────────┤
│  1. Task Context                                        │
│     "You are an AI assistant that..."                   │
├─────────────────────────────────────────────────────────┤
│  2. Tone Context                                        │
│     "Be factual, confident, avoid guessing..."          │
├─────────────────────────────────────────────────────────┤
│  3. Background Data                                     │
│     Static info: form structure, field meanings, etc.   │
└─────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────┐
│                      USER PROMPT                         │
├─────────────────────────────────────────────────────────┤
│  4. Dynamic Content                                     │
│     The actual input (images, text, data)               │
├─────────────────────────────────────────────────────────┤
│  5. Detailed Instructions                               │
│     Step-by-step task breakdown                         │
├─────────────────────────────────────────────────────────┤
│  6. Examples (Few-Shot)                                 │
│     Input/output pairs demonstrating behavior           │
├─────────────────────────────────────────────────────────┤
│  7. Output Format                                       │
│     JSON, XML, specific structure                       │
├─────────────────────────────────────────────────────────┤
│  8. Task Reminder                                       │
│     "Remember to..."                                    │
├─────────────────────────────────────────────────────────┤
│  9. Important Guidelines                                │
│     Critical constraints and rules                      │
└─────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────┐
│                   ASSISTANT PREFILL                      │
├─────────────────────────────────────────────────────────┤
│  10. Prefill                                            │
│      Start response: "{" or "<result>"                  │
└─────────────────────────────────────────────────────────┘

When to Use Each Element

Element	Use When...	Skip When...
Task Context	Always	Never
Tone Context	Specific voice needed	Default tone is fine
Background	Static info exists	Info is in input
Dynamic Content	Processing data	Pure generation
Instructions	Complex task	Simple task
Examples	Precision critical	Clear enough
Output Format	Parsing needed	Free-form OK
Reminder	Long prompt	Short prompt
Guidelines	Edge cases exist	Simple task
Prefill	JSON/XML output	Free text

Example: Simple vs. Structured Prompt

Simple (Unreliable)

Summarize this article.

Structured (Reliable)

Task: You are a content summarizer that creates concise summaries.

Tone: Be factual and neutral. Avoid editorializing.

<article>
{article_content}
</article>

Instructions:
1. Read the entire article
2. Identify the main thesis
3. Extract 3-5 key points
4. Write a 2-3 sentence summary

Output Format:
<summary>
<main_point>...</main_point>
<key_points>
  <point>...</point>
</key_points>
<summary_text>...</summary_text>
</summary>

Remember: Only include facts from the article. Do not add external information.

System Prompt vs User Prompt

System Prompt (via API):

Persistent across conversation
Contains static context
Defines role and behavior
Great for caching

User Prompt:

Changes per request
Contains dynamic input
Includes specific task details

Order Matters

Research shows that information placement affects results:

Put critical info at start and end - The "primacy" and "recency" effects
Group related information - Easier to reference
Progressive disclosure - Context before details

Key Takeaways

Structure matters - Organized prompts produce better results
Use the 10-point framework - As a checklist, not a requirement
Separate system and user prompts - Static vs. dynamic content
Order information strategically - Important info at start and end
Not all elements needed - Use what's relevant to your task

Next Lesson

Next, we'll explore Task and Tone Context - the first two critical elements that set the stage for everything else.

Launch Offer