7 Prompt engineering: Strategies for guiding and evaluating LLMs

Prompt engineering is presented as a practical discipline for shaping large language model behavior by translating broad intent into clear, contextualized instructions. The chapter explains how small changes in wording, audience, and format can dramatically affect results, and why providing the right context—via system instructions, examples, constraints, and retrieved information—remains essential even as models improve. It highlights both the power and the limits of prompting: models are stateless, strategies can be brittle across versions, and some tasks demand multistep workflows and tool use. As the field matures, prompt engineering broadens into context engineering, which orchestrates system prompts, conversation history, retrieval, tools, and output schemas to align model behavior with user goals.

Building on core principles—clarity, relevant context, structure, and iteration—the chapter surveys foundational techniques (zero-shot, few-shot, multi-turn) and scaffolds for reasoning and planning (chain-of-thought, tree-of-thought, self-consistency), quality control (reflexion and verification), task decomposition (least-to-most, decomposition, generated knowledge, pyramid priming), modular pipelines (prompt chaining), role cues, and explicit context management. It then introduces higher-level frameworks that make prompts more consistent and reusable: ReAct integrates reasoning with tool calls; MECE enforces clean, comprehensive categorization; PACT encodes persona, action, context, and tone; and WISER structures who, instruction, subtasks, examples, and review. Evolving practices emphasize iterative design and measurement, prompt libraries and versioning, multimodal prompting, adaptive clarification dialogs, and multi-agent orchestration that critiques, votes, and synthesizes for reliability.

A central theme is evaluation: define success criteria, assemble representative datasets, and score outputs with a mix of deterministic checks, expert labels, LLM-as-judge prompts validated against humans, and production A/B testing. The chapter warns about evaluator bias and reward hacking, and recommends comprehensive regression testing to prevent unintended trade-offs when prompts or models change. It contrasts prompting—fast, accessible, and cost-effective but sometimes blunt—with post-training, which enables precise behavioral control and can distill behaviors into smaller, cheaper models once good prompts and datasets exist. The guidance is to start with strong prompts and systematic evaluation, then apply post-training where nuance, robustness, or efficiency demands it, all while adhering to the enduring principles of clarity, context, structure, and iterative improvement.

Prompt engineering works by crafting inputs (prompts) that guide the model’s output.

Structure of a prompt combining system role, context, and user input

An example annotation prompt to judge real estate listing descriptions

• Prompt engineering, once a niche academic interest, has become a full-fledged discipline and design process with the advent of widespread generative AI applications.
• Beyond standard prompting techniques, such as zero-shot, few-shot, and multi-turn prompting, advanced techniques (chain-of-thought, tree-of-thought, and more) can further improve model responses by prompting the model to produce reasoning chains or multiple possible solutions, to reflect on them, and to verify its response.
• Prompting frameworks, including ReAct, MECE, PACT, and WISER, provide simple and reusable recipes for producing structured model outputs.
• To effectively measure and iterate on prompts, we need to develop evaluations that include sample contexts and methods of scoring the model responses produced by each prompt.
• Although prompting and prompt engineering are crucial to controlling model outputs, in some cases, post-training or a hybrid method may be more appropriate.

FAQ