Prompt engineering

What prompt engineering is

Prompt engineering is the practice of designing the input text that steers a generative AI system toward a desired output ^[1]. In a generative system, the user does not usually program every step of the computation; instead, the prompt provides instructions, examples or constraints that shape what the model generates next. The model’s parameters are not permanently changed by a prompt, but the wording, examples and structure of the input can change the response. Prompt design is therefore a direct interface between people and generative models.

Chain-of-thought as a worked example

One widely discussed technique is chain-of-thought prompting, introduced by Wei et al. in January 2022 ^[2]. The prompt supplies a few demonstrations that show intermediate reasoning steps before the final answer ^[2]. Wei et al. reported that chain-of-thought prompting improves performance across arithmetic, commonsense and symbolic reasoning tasks ^[2]. In one result, prompting a 540B-parameter model with just eight chain-of-thought exemplars reached state-of-the-art accuracy on the GSM8K math word problem benchmark, surpassing even a fine-tuned GPT-3 with a verifier ^[2]. The authors also found that this reasoning ability emerges naturally only in sufficiently large language models ^[2].

A fragmented field

The Prompt Report, published in June 2024, notes that prompt engineering suffers from conflicting terminology and a fragmented understanding of what makes an effective prompt ^[1]. To address this, the survey presents a vocabulary of 33 terms and a taxonomy of 58 LLM prompting techniques and 40 techniques for other modalities ^[1]. It also provides best practices for prompting state-of-the-art models and a meta-analysis of the natural language prefix-prompting literature ^[1].

Why shared vocabulary matters

Terminology matters because techniques are named and described inconsistently across papers; without a shared vocabulary, results are hard to reproduce and compare ^[1]. Chain-of-thought prompting illustrates another limit: the reasoning ability it draws on emerges naturally only in sufficiently large language models, so a technique that works for one model may not transfer unchanged to a much smaller one ^[2]. The Prompt Report’s vocabulary and taxonomy are an attempt to make claims about prompting more precise and comparable ^[1].

Where it is used

Prompt engineering sits at the point where people interact with generative AI. It is used to specify tasks, provide examples and shape outputs. The Prompt Report covers both LLM prompting and techniques for other modalities, indicating that prompt design is not limited to text-only models ^[1]. At the same time, inconsistent terminology and fragmented understanding remain obstacles to comparing and reproducing prompting methods ^[1].