A structured map of the techniques used to steer a generative AI model purely through the text of its input, without changing model weights. It covers the shared vocabulary (prompt, prompt template, exemplar, prompt chain, answer engineering) and organizes text-only techniques into families: in-context learning and the design decisions behind few-shot exemplars (quantity, ordering, label distribution, format, similarity), zero-shot techniques such as role and emotion prompting, thought generation including Chain-of-Thought, decomposition such as Least-to-Most and Tree-of-Thought, ensembling such as Self-Consistency, and self-criticism such as Self-Refine and Chain-of-Verification. It extends to multilingual and multimodal prompting, prompts that drive agents which call tools or retrieve documents, prompts that make a model act as an evaluator, and the failure surface of prompting: injection, jailbreaking, prompt sensitivity, overconfidence, sycophancy and ambiguity. The scope is deliberately restricted to hard (discrete) prefix prompts, excluding soft prompts and any method that updates weights, so the whole subject is about what can be achieved by rewriting text alone.
A systematic review of prompt engineering, the iterative practice of rewriting the text input given to a generative AI model in order to improve its output. Scope is restricted to hard (discrete) pre…