appendix-prompt-engineering

Appendix A - Prompt Engineering

Overview

Prompt Engineering is the disciplined practice of crafting, iterating, and optimizing the inputs given to language models to produce reliable, high-quality outputs. It is not a simple act of asking questions — it is a structured engineering discipline that transforms a general-purpose language model into a specialized, highly capable tool for specific tasks.

For agents, prompting is the foundational layer: every agentic pattern depends on well-crafted prompts. A powerful model with a poor prompt produces poor results. A well-engineered prompt transforms model outputs from probabilistic guesses into deterministic, structured, trustworthy cognitive operations.

Core Principle: Treat prompts as code — version them, test them, iterate them, and document what works and why.

When This Skill Applies

Activate this pattern when:

• Building any LLM-powered agent that needs reliable, structured outputs
• Writing system prompts, user instructions, or tool descriptions for agents
• An agent produces inconsistent, hallucinated, or poorly formatted responses
• You need structured data (JSON, XML) from model outputs
• Implementing Chain-of-Thought, ReAct, or few-shot reasoning patterns
• Testing and comparing different prompt strategies
• Preparing prompts for production deployment

Rule of thumb: Every agent interaction is a prompt. Engineer them deliberately — don't leave agent behavior to chance.

Prompting Technique Hierarchy

Zero-Shot: No examples — just instructions
    → Simple tasks with clear instructions

One-Shot: One example — demonstrate the pattern
    → When output format matters

Few-Shot: 2–5 examples — show the range
    → Complex tasks, classification, structured extraction

Chain-of-Thought: "Think step by step" + examples
    → Reasoning tasks, math, multi-step problems

ReAct: Thought → Action → Observation loop
    → Agents using tools in an iterative loop

DEFINE → PLAN → ACTION Workflow