What Is Prompt Engineering?

Prompt engineering is the practice of designing the inputs to an AI model to get better outputs.

LLMs are sensitive to how questions are framed. The same model, given two differently worded inputs on the same topic, can produce outputs that range from excellent to useless. Prompt engineering is the discipline of understanding why this happens and using that understanding to get consistent, reliable results.

Why Prompts Matter

A large language model does not have a fixed set of rules it follows. It generates responses based on patterns learned during training, shaped by the specific words and context you give it. This means:

• Ambiguous instructions produce inconsistent outputs. "Write a summary" is different from "Write a 3-bullet executive summary of the key decisions and open questions, for an audience of senior stakeholders who have not read the document."
• Context changes outputs. Including relevant background information, examples, or constraints shapes what the model produces — and what it leaves out.
• Format specification matters. If you need structured output (JSON, a table, bullet points), specify it explicitly. Models will default to natural prose otherwise.

Core Techniques

Zero-shot prompting — asking the model to do something directly, with no examples. Works for simple, clear tasks. Unreliable for complex or ambiguous ones.

Few-shot prompting — including examples of the desired input/output pattern before the actual request. Dramatically improves consistency on pattern-following tasks: classification, extraction, formatting.

Chain-of-thought prompting — instructing the model to reason step-by-step before giving an answer ("Think through this step by step before responding"). Improves accuracy on reasoning and logic tasks.

System prompts — instructions provided before the user's message that define the model's role, constraints, and behaviour for the entire session. Central to all production AI applications. Example: "You are an email triage assistant. Classify each email as: enquiry, complaint, or other. Never respond to the sender. Always output JSON."

Output format specification — explicitly defining what structure the response should take. JSON output is the standard for AI applications where the output needs to be parsed by code downstream.

Negative instructions — telling the model explicitly what not to do. "Do not make up information you are not certain about" is more reliable than expecting the model to avoid hallucination without being told.

Prompt Engineering in Practice

For a production business application — an AI that reads incoming emails and classifies them, for example — a good system prompt addresses:

• Role: What is the AI's job?
• Input: What will it receive?
• Task: What should it do with that input?
• Output format: Exactly what should the response look like?
• Constraints: What should it avoid? What should it do when uncertain?
• Examples: (Optional) A few examples of input and expected output

Getting these right makes the difference between an AI agent that works reliably and one that fails unpredictably on real-world inputs.

The Broader Picture

Prompt engineering does not replace fine-tuning or RAG (retrieval-augmented generation) for every use case. For tasks requiring very specific knowledge or behaviour:

• RAG gives the model relevant information it was not trained on (documents, your knowledge base)
• Fine-tuning trains the model on domain-specific examples to shift its baseline behaviour

But prompt engineering is always the starting point — and for most business applications, a well-crafted prompt is all that is needed.

---

WhatWill AI builds AI agents and automation systems with robust prompt engineering at the core. Book a discovery call to discuss what we can build for your business.