Skip to main content

Prompt Engineering Guide

Prompt engineering is the art of designing inputs that get the best outputs from large language models. This guide covers practical techniques for production use.

Core Techniques​

1. Zero-Shot Prompting​

Give the model a task with no examples:

Classify the following support ticket as one of: bug, feature_request, question, billing

Ticket: "The dashboard keeps crashing when I filter by date range. This started after the latest update."

Classification:

2. Few-Shot Prompting​

Provide examples to guide the model:

Classify support tickets:

Ticket: "How do I reset my password?" → question
Ticket: "Add dark mode to the settings page" → feature_request
Ticket: "I was charged twice this month" → billing
Ticket: "The export button returns a 500 error" → bug

Ticket: "Can you add Slack integration?" →

3. Chain-of-Thought (CoT)​

Ask the model to reason step by step:

A Kubernetes cluster has 3 nodes, each with 8 CPUs and 32GB RAM.
Each pod requests 0.5 CPU and 512MB RAM.
The system components use 1 CPU and 2GB RAM per node.

How many pods can the cluster run? Think step by step.

4. Role-Based Prompting​

You are a senior SRE with 10 years of experience in Kubernetes operations.
A team is experiencing intermittent 503 errors on their service.
Their HPA is configured with a target of 80% CPU.
Current pod count: 3, CPU usage: 92%.

Diagnose the issue and provide a remediation plan.

Production Prompt Patterns​

Structured Output​

Analyze this Terraform plan and respond in JSON format:

{
  "risk_level": "low|medium|high|critical",
  "resources_added": <number>,
  "resources_changed": <number>,
  "resources_destroyed": <number>,
  "concerns": ["list of potential issues"],
  "recommendation": "approve|review|reject"
}

Terraform plan:
...

System + User Message Pattern​

messages = [
    {
        "role": "system",
        "content": """You are an AIOps assistant that analyzes 
        infrastructure metrics. You respond with:
        1. Current status assessment
        2. Anomalies detected
        3. Recommended actions
        Keep responses concise and actionable."""
    },
    {
        "role": "user",
        "content": f"Analyze these metrics:\n{metrics_data}"
    }
]

Template with Variables​

INCIDENT_ANALYSIS_PROMPT = """
Analyze this incident and provide a structured post-mortem:

Service: {service_name}
Alert: {alert_name}
Duration: {duration}
Impact: {impact}
Timeline: {timeline}

Provide:
1. Root Cause Analysis
2. Contributing Factors
3. Action Items (with owners and deadlines)
4. Lessons Learned
"""

Prompt Engineering Best Practices​

PracticeExample
Be specific"List 5 items" not "List some items"
Set constraints"Respond in under 100 words"
Define format"Use bullet points" or "Respond in JSON"
Provide contextInclude relevant background information
Use delimitersWrap input data in ``` or ---
IterateTest, evaluate, and refine prompts

Common Pitfalls​

  1. Vague instructions — "Make it better" vs. "Improve readability by using shorter sentences"
  2. Missing context — Always provide relevant domain knowledge
  3. No output format — Specify exactly how you want the response structured
  4. Ignoring edge cases — Test with unusual inputs
  5. Prompt injection vulnerability — Sanitize user inputs in production

Next Steps​