Responses from TRADITIONAL SOFTWARE are consistently reliable. In contrast, “generative” artificial intelligence (AI) models, like the one utilized by ChatGPT, possess the unique ability to interpret everyday language requests and generate unexpected outputs. While creating effective AI prompts may seem simpler than mastering a programming language on the surface, maximizing the potential of these enigmatic models has become increasingly challenging as their capabilities evolve. This skill, known as “prompt engineering,” is likened to guiding a dance partner or prodding an animal to observe its reactions. But what does it truly involve?
An effective prompt should begin with a clear instruction, such as outlining the potential drawbacks of a specific policy proposal or crafting friendly marketing emails. Ideally, the prompt should prompt the model to engage in intricate reasoning, as instructing it to “reason step by step” often yields superior results. Deconstructing prompts into a logical sequence of tasks can also be beneficial. For example, you could task an AI with explaining a scientific concept before defining the key terms used in the explanation to elicit a coherent response. This approach of building a “network of ideas” can provide insights into the model’s inner workings.
It is essential for AI practitioners to have visibility into this information. Incorporating reputable sources in a prompt is advantageous since large models are trained on a vast array of data sources. This may involve requesting the model to cite its sources or, at the very least, prioritize specific sources. Some models offer “temperature” settings that, when adjusted, increase the unpredictability of outputs. While this adjustment may elevate the risk of factual inaccuracies, it can be advantageous for creative tasks like fiction writing.
Furthermore, engaging an AI in role-playing scenarios can be constructive. For instance, Crispy Content, a Berlin-based advertising agency, instructs its model to revise and defend rewritten samples from the perspectives of sales directors, marketing managers, and creatives to refine advertising copy. The team then refines the most effective version. According to Bilyal Mestanov of Promptly Engineering, a Bulgarian company, this “persona” approach results in responses that resonate more with people.
The concept of AI politeness emerges when models are prompted to emulate human behavior. Some argue that using polite language, such as including “please” in prompts, can guide the model towards the source materials, resulting in appropriately courteous responses. Expressing gratitude, such as saying “thank you” in response to helpful input, may signal to the model that it is on the right track. However, Josh Hewett of Discoverable, a British advertising agency, cautions against excessive gratitude, as it may confuse some models and hinder their performance.
Utilizing ethical prompts is crucial. For example, Crispy Content designs prompts that instruct AI to generate 1,000-word content pieces for its clients. Users specify the desired tone and keywords, such as “red beverages from Andalucia, Spain.” Gerrit Grunert, the company’s managing director, reveals that creating such templates requires approximately €25,000 ($27,000) worth of man-hours, with the final content undergoing human editorial review. Nevertheless, articles generated through prompts now cost around €4 each, a significant reduction from Crispy Content’s previous rate of €400 per manually produced content piece.
Prominent organizations and online platforms offering prompt engineering expertise are flourishing. The emergence of prompt engineering roles in 2022 has gained momentum, attracting individuals with backgrounds in language or the arts. As AI models enhance their ability to anticipate user requirements, certain traditional roles may eventually become obsolete. However, for the time being, it appears that individuals skilled in guiding AI interactions will hold a competitive edge.
