Table 1. Summary of Several Prompt Engineering Methods in LLMs.

Prompt Engineering	Principle	Features	Applications
Zero-shot	Directly provides task description without examples.	Simple to use, no additional data needed.	Simple classification, generation tasks (e.g., text mining of MOF synthesis27)
Few-shot	Provides a few examples to guide the model.	Improves model understanding of the task.	Moderately complex tasks (e.g., property prediction through SMILES21)
CoT	Guides the model to reason step-by-step.	Suitable for complex reasoning tasks.	Math problems, logical reasoning (e.g., calculating chemical equilibrium constants45)
APE	Automatically generates and optimizes prompts using the model’s own capabilities.	Reduces manual effort; may produce more effective prompts than human-designed ones.	Tasks requiring efficient prompt design.
ReAct	Solves tasks through dynamic reasoning and external actions	Suitable for multistep reasoning and external interaction tasks; improves transparency.	Complex question answering, tasks requiring external knowledge (e.g., prediction and generation of MOFs50)
RAG	Combines retrieval from external knowledge bases with generation to produce accurate answers.	Improves accuracy and reliability; handles tasks requiring external knowledge.	Open-domain question answering, fact-based tasks (e.g., transform words in battery research13)
Meta-prompting	Uses a meta-prompt to guide the model in generating specific subprompts or task decompositions.	Enhances model’s ability to understand and execute complex tasks; highly flexible.	Complex task decomposition, multistep reasoning tasks (e.g., autonomous chemical research1)