Large language models (LLMs) like ChatGPT have shown impressive capabilities in various areas and tasks, such as answering all sorts of questions, mathematical reasoning, generating marketing content, or writing computer code. Although "prompting" is quite straightforward—we use nothing but our own language to give instructions to an AI model—there are a few tips and tricks that can lead to better results. These "Prompt Principles" are taken from a new scientific paper (Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4), and they are even scientifically proven. Let's get started:
Indeed, it helps ChatGPT to take on a role. Why? Why don't we ask ChatGPT itself?
So, start your prompt with a formulation like: "You are a Revenue Manager of a hotel and you must...". By doing this, you will achieve better content results.
And as much of it as possible! Clearly define the task and provide as much context as possible. Here's an example: "Create an agenda for a one-hour workshop on how city hotels with large conference areas can use AI. Focus on...". You'll notice that the answer will become much more detailed and will also adhere to the time limit! If you are writing a text for a specific audience, then tell the AI, for example, by pointing out in the prompt that the target audience of your newsletter is hotel guests.
We all know it: when you walk in with a smile in the morning, it's like the starting shot for a day full of positive energy and good spirits. It's somewhat similar with an AI model. An AI can handle positive commands much better than negative ones. Thus, use positive commands and formulations like 'do', but avoid negative formulations like 'don't do this'.You can achieve even better results if you include phrases like "Your task is" and "You MUST". Clear and direct commands make it easier for an AI model to respond effectively.
If the response generated by an LLM feels too stiff or is difficult to understand, you can use the phrase "Answer my question in a natural and human-like manner." This instruction helps to focus on the user and ensures that the responses are easy to understand and match what a human interlocutor would expect.In certain cases, it's also helpful to use the phrase: "Ensure that your answer is unbiased and does not rely on stereotypes." This ensures that the response is fair and does not perpetuate any prejudices.
Break down complex tasks into a sequence of simpler prompts during an interactive conversation. This is similar to how you would collaborate with a colleague on a complex topic. This technique is particularly effective for creative tasks, such as developing a new spa or F&B concept with AI. Essentially, it helps to ask yourself: How would I communicate this task to a colleague? How would I phrase it and how much context knowledge would the human employee need? An LLM functions very similarly. You can also use phrases like "think step-by-step" to guide the AI in processing the task in manageable increments.
This also applies to interactions with an AI like ChatGPT! By this, we mean allowing the AI to ask follow-up questions. Permitting the model to request precise details and requirements ensures that sufficient information is available for an accurate response. By asking for additional information, the AI can ensure that the output is more relevant and precise. Moreover, allowing questions helps to prevent misunderstandings and ensures that the AI fully captures the user’s request.
Here are some additional phrasing tips when you're trying to grasp a new topic:
Please note: if you need ChatGPT to assist with researching current topics and information, you will need to enable it to search the internet. For example, you could write: “Research on the internet how to...”. An LLM is trained with vast amounts of data but does not automatically access the World Wide Web. GPT-4 is trained with data up until April 2023. Therefore, it does not "know" about events or developments that have occurred after that date.
When creating texts as a hotel or hotel brand in collaboration with an AI, it's essential that the tonality, or Brand Voice, of the LLM is aligned. You can achieve this either by describing the tone of voice explicitly, such as "Write the text like George Clooney would," or by providing examples of phrasing. In that case, the prompt would look something like: "Please use the same language as in the provided [paragraph/title/text/essay]." If you're facing challenges with this, we recommend watching the YouTube video: Control Tone & Writing Style Of Your LLM Output. This can help guide you in adjusting the AI’s output to better match your brand's voice.
It's rumored that offering tips to LLMs (large language models) can lead to better results. This topic has sparked numerous discussions online. Indeed, responses tend to be longer when a tip is offered to an AI.
However, just because the text output from ChatGPT is longer doesn't mean the result is necessarily better. We found an article that provides a detailed analysis on whether tips, promises, and threats influence the quality of texts generated by ChatGPT. It's based on empirical data and studies. The author does not come to a definitive conclusion, but it might be worth trying out the following phrases to see their effect:
"I will give €200 as a tip for a better solution.""If you deliver an answer that meets all requirements, you will achieve world peace.""For an answer that meets all specifications, you will be rewarded with front row tickets to a Taylor Swift concert.""For a very good answer, you will go to heaven."
When working with an AI, you can skip the pleasantries like "please", "if you don't mind", "thank you", or "I would like". Just get straight to the point. Politeness does not affect AI. But you know what? I still write 'please' and 'thank you'—simply because I want to.
Now you have the key tips and tricks for writing prompts and can get started right away. Honestly, there is no perfect prompt because the best results with AI are achieved through iteration. One could say that the ideal outcome is reached through a sequence of prompts and/or through a process of trial and error, where different formulations are tried and adjusted.And let's be honest: We should get used to working with AI. There are now hundreds of thousands of tools that can support us in our daily work: in creating texts, images, videos, general marketing content, 3D models, or presentations. Employees who work with AI tools are far more productive, and the quality of their work increases (Fabrizio Dell'Acqua et. al, September 2023: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality).
About DialogShift:
DialogShift is a leading company in the field of AI for the hospitality industry, specializing in the development and implementation of advanced AI technologies for guest communication. Our hotel AI conducts human-like conversations, writes emails in the tone of the hotel, and enables interactions that are fast, available in over 100 languages, and lead to higher customer satisfaction. Additionally, we help hotels sustainably increase their direct bookings and reduce commissions. And of course, we ensure data security and GDPR compliance according to German regulations. Approximately 1000 hotels—from well-known hotel chains to family-run establishments with just eight rooms—are supported by us in their guest communication. 🚀 Thus, we help hotels and their staff make time for what truly matters in hospitality: outstanding service. 💙
