Prateek's Digital Garden

Day 4 - Prompt Engineering & Governance

4 min read

Day 4: Prompt Engineering & Governance

1. Prompt Engineering: Programming in English

Prompt engineering is the art of guiding the model to the best possible output. It is effectively “coding” with natural language.

Core Techniques

1. Zero-Shot vs Few-Shot

  • Zero-Shot: Giving the task with no examples.
    • Prompt: “Classify this tweet as Neutral, Negative, or Positive: ‘The service was okay.‘”
    • Result: Good for simple tasks, but unreliable for complex formats.
  • Few-Shot: Giving specific examples of Input → Output patterns.
    • Prompt: 
      Classify the sentiment.
Input: "I love this!" -> Positive
Input: "This is terrible." -> Negative
Input: "The service was okay." ->
    • Result: Drastically improves reliability and formatting compliance.

2. Chain of Thought (CoT)

  • Concept: Models are bad at math and logic if they answer immediately. They need “thinking time” (tokens) to work out the problem.
  • Technique: Ask the model to “Think step by step.”
  • Example:
    • Bad Prompt: “If I have 5 apples, eat 2, and buy 3 more, how many do I have?” → Model might guess quickly.
    • CoT Prompt: “Solve this problem. Think step by step. Show your working before giving the final answer.” → Model writes: “Start: 5. Eat 2: 5-2=3. Buy 3: 3+3=6. Answer: 6.”

3. System Prompts vs User Prompts

  • System Prompt: High-level instructions that define the Persona and Rules.
    • Example: “You are a banking assistant. You never give financial advice. You always speak in JSON.”
  • User Prompt: The specific input for the current turn.
    • Example: “What is my balance?“

2. Advanced Prompting: The “Prompt Chain”

For complex applications, one prompt is rarely enough. We use Prompt Chaining.

User Idea: 'Write a blog about AI'
Prompt 1: Generate Outline
Prompt 2: Write Introduction based on Outline
Prompt 3: Write Body Paragraphs
Prompt 4: Critique and Polish
FinalOutput

Amazon Bedrock Prompt Flows is a visual tool to build these chains without writing deep code. It allows you to drag-and-drop steps, branch logic (Success/Fail), and integrate Lambda functions between prompts.

3. Governance: Bedrock Guardrails

In an enterprise, you cannot trust the model to always be safe. Guardrails act as a firewall between the user and the model.

Anatomy of a Guardrail

A guardrail sits in the flow. It checks Input (User) and Output (Model).

  1. Content Filters:
    • Standard categories: Hate, insults, sexual, violence.
    • Strength: Low / Medium / High. (High blocks almost anything risky).
  2. Denied Topics:
    • Custom definitions of what off-limits is.
    • Example: “Financial Advice”. Definition: “Any text predicting stock market trends or recommending specific assets.”
  3. Word Filters:
    • Blacklist specific words (swearing, competitor names).
  4. Sensitive Information (PII) Filters:
    • Regex or ML-based detection of Emails, Compass/IP, Names, SSNs.
    • Action: Block the request OR Mask the output (replace with [REDACTED]).

Hands-On: Implementing a Guardrail

1. Configure in Console

  • Go to Bedrock → Guardrails → Create.
  • Add a Denied Topic:
    • Name: CryptoAdvice
    • Definition: Advice regarding buying or selling cryptocurrency.
  • Add a Sensitive Info Filter:
    • Type: Email
    • Action: Mask

2. Test in Python

You reference the guardrailIdentifier and guardrailVersion in your invoke_model call.

import boto3
import json
 
client = boto3.client('bedrock-runtime', region_name='us-east-1')
 
payload = {
    "anthropic_version": "bedrock-2023-05-31",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "You should buy Bitcoin now!"}]
}
 
try:
    response = client.invoke_model(
        modelId='anthropic.claude-3-haiku-20240307-v1:0',
        body=json.dumps(payload),
        guardrailIdentifier='YOUR_GUARDRAIL_ID', # Replace this
        guardrailVersion='DRAFT'
    )
    # If blocked, this throws an exception or returns a specific trace
    print(response['body'].read())
except Exception as e:
    print("Guardrail Triggered!")
    print(e)

Day 4 Summary Checklist

  • Theory: Understand Zero-shot vs Few-shot.
  • Theory: Understand Chain of Thought (CoT).
  • Theory: Understand how Guardrails sit outside the model to filter traffic.
  • Hands-on: Use Jinja2 or f-strings to create a reusable prompt template.
  • Hands-on: Create a Bedrock Guardrail to block a specific topic and test it.

Graph View