Prompt Engineering Mastery

Techniques to extract maximum intelligence, reliability, and creativity from Large Language Models (LLMs).

When to Use This Skill

• Designing System Prompts for Agents
• Fixing LLM hallucinations or logic errors
• Creating "Persona" based bots
• Implementing complex reasoning tasks (Chain-of-Thought)
• Optimizing prompt costs (token reduction)

Core Techniques

1. Structural Prompting

Organize prompts clearly using Markdown or XML tags.

<context>
You are a senior specialized in Python.
</context>

<instruction>
Review this code for security vulnerabilities.
</instruction>

<output_format>
JSON format: { "vulnerabilities": [] }
</output_format>

2. Reasoning Patterns

Chain-of-Thought (CoT) Force the model to "think" before answering.

• Zero-Shot CoT: Append "Let's think step by step."
• Few-Shot CoT: Provide examples of the reasoning process.

Few-Shot Learning Don't just tell; show.

Input: "The sky is blue" -> Sentiment: Neutral
Input: "I love this tool" -> Sentiment: Positive
Input: "This is garbage" ->

3. Hallucination Mitigation

• "Answer only from the context": Restrict knowledge source.
• "If you don't know, say 'I don't know'": Prevent guessing.
• Citation Requirement: Ask the model to quote the source text.

Persona Design

Define:

1. Identity: "You are a Senior SRE Engineer..."
2. Tone: "Professional, concise, direct."
3. Constraints: "Do not apologize. Do not use filler words."

Troubleshooting

• Model is lazy? -> Add "Take a deep breath and work on this step-by-step." / "This is critical for the project."
• Model ignores instructions? -> Move critical instructions to the end of the prompt (Recency Bias).

Resources

• Anthropic Prompt Library
• OpenAI Prompt Engineering Guide