Prompt Engineering

Definition

Prompt engineering is the practice of designing and refining the input text (prompt) given to a language model to reliably elicit accurate, relevant, and well-formatted outputs. It involves techniques like providing clear instructions, examples (few-shot prompting), structured formats, and reasoning scaffolds to steer a model's behavior without changing its underlying weights.

Overview

Large language models generate output based entirely on the text they are given as context, so how a request is phrased significantly affects the quality of the response. Prompt engineering emerged as a discipline for systematically crafting these inputs to get consistent, high-quality results, particularly important in production applications where reliability matters more than a single lucky output. Common techniques include zero-shot prompting (just asking directly), few-shot prompting (providing example input-output pairs to demonstrate the desired pattern), and chain-of-thought prompting (asking the model to reason step by step before answering, which improves accuracy on complex tasks). Other techniques include role prompting (asking the model to adopt a persona or expertise), specifying output format explicitly (e.g., JSON schemas), and providing negative constraints (what not to do). Prompt engineering also encompasses system-level prompt design in applications: crafting a 'system prompt' that sets persistent behavior and tone for a conversation, structuring multi-turn interactions, and combining retrieval-augmented generation (RAG) to inject relevant external context into the prompt so the model can answer questions beyond its training data. As models have become more capable, some early prompting tricks have become less necessary, but the discipline remains essential for building reliable applications: reducing hallucination, controlling tone and format, minimizing token cost, and adapting a general-purpose model to a specific domain without the expense of fine-tuning. Prompt engineering is often the fastest and cheapest lever for improving an LLM application's behavior before considering more expensive options like fine-tuning or building custom retrieval pipelines.

Key Concepts

• Shapes model output purely through input text, without modifying model weights
• Includes zero-shot, few-shot, and chain-of-thought prompting techniques
• System prompts set persistent behavior, tone, and constraints for an application
• Output format can be explicitly constrained (e.g., JSON, XML, specific templates)
• Iterative and empirical: prompts are tested and refined based on observed outputs
• Often the cheapest and fastest lever for improving LLM application quality
• Complements, rather than replaces, techniques like RAG and fine-tuning
• Model- and version-specific; effective prompts can behave differently across models

Use Cases

Frequently Asked Questions

Related in AI Fundamentals