Separating the Brain from the Meat

In the early stages of development, you might just send one big string to the AI. But in a professional application, you must distinguish between the System Prompt (the instructions) and the User Prompt (the payload).

In the AWS Certified Generative AI Developer – Professional exam, Domain 4 tests your ability to manage these inputs as separate architectural assets.

---

1. The System Prompt (The Foundation)

The System Prompt is the "Constitutional Law" of your AI agent. It is set by the developer and the user usually never sees it.

• Persona: "You are a professional accountant."
• Rules: "Never discuss politics. Always respond in JSON."
• Knowledge Boundaries: "Only answer questions about the provided context."

Why separate it? By keeping instructions in the System Prompt, you reduce the risk of a user "overwriting" the rules through prompt injection.

---

2. The User Prompt (The Variable)

The User Prompt is the raw, unvalidated input from the customer.

• Example: "What is my current bank balance?"

The Pro Path: Never send the raw user prompt directly to the model. Always wrap it in a structure that provides context.

---

3. Dynamic Prompt Templates

Your code shouldn't be hardcoding text. It should use Templates.

# A professional prompt template strategy
PROMPT_TEMPLATE = """
You are a support agent for `{company_name}`.
Below is the customer's history:
`{customer_history}`

Customer Question: `{user_input}`
"""

# At runtime:
final_prompt = PROMPT_TEMPLATE.format(
    company_name="Cloud Corp",
    customer_history="Ordered 1 widget in 2023.",
    user_input="Where is my order?"
)

---

4. Amazon Bedrock Prompt Management

AWS recently released a managed service to replace hardcoded strings: Amazon Bedrock Prompt Management.

• Versioning: You can create v1, v2, and v3 of a prompt and switch between them without redeploying your Lambda code.
• Testing: Test your prompts directly in the AWS console against multiple models.
• Parameterization: Define {user_name} or {context} placeholders in the console, and Bedrock will fill them in when you call the API.

graph LR
    App[Your App] --> B[Bedrock Prompt API]
    B -->|Get Prompt v5| Registry[Prompt Registry]
    Registry -->|Return Prompt| App
    App -->|Send to Model| BedrockRun[Bedrock Runtime]

---

5. Security: The "Separator" Pattern

Attacks like "Prompt Injection" occur when the model gets confused about which part of the text is an "Instruction" and which part is just "Data." The Pro Solution: Use clear delimiters to separate them.

• ### INSTRUCTIONS ###
• ### USER INPUT ###
• [CONTEXT_START] / [CONTEXT_END]

This helps the model's self-attention mechanism focus on the right parts of the prompt at the right time.

---

6. Token Budgeting

Prompt management is also about Economics.

• Every word you add to your System Prompt (e.g., repeating "Be polite" five times) costs you money on every single request.
• The Optimization: Periodically "Audit" your prompts to remove redundant instructions or filler words that don't change the model's output quality.

---

Knowledge Check: Test Your Prompt Management Knowledge

---

Summary

Don't treat prompts like strings; treat them like Configurations. By separating System and User inputs and using managed registries, you build a system that is secure, updateable, and cost-efficient. In the next lesson, we will look at Optimizing Prompts for Different Models.

---

Next Lesson: Cross-Model Engineering: Optimizing Prompts for Different Models