9 Advanced prompt patterns

The chapter expands the prompt engineer’s toolkit beyond basic structure, templates, examples, and aggregation by introducing higher-level patterns that shape how a model reasons and interacts. The main idea is that some problems are better solved by making reasoning explicit, generating prompts from descriptions, or looping through evidence and action rather than relying on a single direct request.

Chain of thought is presented as a way to expose intermediate reasoning so multistep decisions become auditable and easier to verify, especially when using standard models or when the reasoning logic is specialized to a domain. The chapter shows that direct prompts can produce correct answers without revealing how they were reached, while zero-shot and few-shot variants of chain of thought make the logic visible and, when necessary, teach domain-specific decision rules through a worked example. It also notes that if a reasoning-capable model is already doing internal reasoning, explicit chain-of-thought instructions usually add cost without benefit.

Meta prompting uses a prompt to generate a reusable prompt template, which is useful when a team repeatedly solves the same kind of task and can describe the desired structure clearly. ReAct goes further by combining reasoning with action and observation, making it suitable for investigations where new information must be gathered step by step before drawing a conclusion. Across all three patterns, the chapter emphasizes judgment: use the more advanced pattern when it materially improves reasoning, traceability, or information gathering, and prefer simpler prompting when the task is straightforward or the overhead is not justified.

Advanced pattern selection: comparison of CoT, meta prompting, and ReAct, showing how each pattern differs by reasoning visibility, prompt design requirements, and dependence on iterative interaction with external systems

ReAct loop. The model reasons from current evidence, proposes one action to resolve an information gap, receives the observation as new context, then reasons again. The loop repeats until a conclusion is supported by sufficient evidence.

• CoT makes intermediate reasoning steps explicit and auditable. Use it on nonreasoning models or when the reasoning pattern itself is domain-specific and outside the model’s training distribution.
• Meta prompting uses a prompt to design a prompt template. Applying the structural elements, linguistic elements, and prompt patterns framework as the specification language produces structurally sound templates in one pass.
• ReAct structures agentic workflows as thought-action-observation loops, giving a model a principled way to incorporate real-world observations before committing to a conclusion.
• In pattern selection, use CoT when multistep reasoning must be auditable, meta prompting when prompt design is itself the recurring task, and ReAct when a diagnosis or decision requires iterating on information that cannot all be provided up front.

FAQ