The Four Principles of AI Prompting

In one form or another, AI has crept into many publishing workflows. And as in many human endeavors, there are power AI users and there are beginners. Our intent in this article is to help turn the newbies into ninjas by adopting a few best practices in communicating with AI.

We are talking specifically here about conversing with Large Language Models (LLMs) like ChatGPT, although the number of model types is spreading. Interestingly, this includes “Small” Language Models that are confined to specific applications and built from smaller sets of targeted training data (several billion parameters compared to hundreds of billions of parameters in LLMs) but which work in a comparable way. Among other applications, our company uses Microsoft’s general LLM, Copilot for research to write blogs like this.

What AI is not

LLMs can be used for a myriad of different tasks, in simple jobs like proof-reading, translations and content summarization; however we are focused here on two generic textual applications, one being research and the other for generating creative ideas – for marketing materials, as an example.

In using AI for research, it’s worth reminding our readers that AI or chatbots do not “know” the answers to anything. They merely predict the next sequence of letters and words based on recognizing language patterns, and so might offer reasonable, albeit dull responses to our questions. That is as far as we might want to take it. Otherwise, there is a danger that we give unearned credibility to AI’s responses, which eventually reinforces the notion that what we are reading is fact. It is, of course, important to verify AI’s answers through legitimate and trusted sources, but it is easy to skip. And what are trusted sources? Organizations such as the Associated Press, the BBC and the New York Times were at one time widely regarded as impeccable sources, but today there are large sections of our community that would dispute that. Remember, the word of the year in 2018 according to dictionary.com was “misinformation.” So beware!

A Well-Educated Teenager

We should treat AI as an extremely well-educated but mischievous teenager with a large vocabulary and a fertile mind. One day they will grow up but for now, be careful how and what you ask them! However, there are ways to get the best out of them. Here are four principles that can guide us on how to prompt AI for best results, adapted from Microsoft’s guide on the subject. Generally, the more detailed the question, the better the answer. These principles are:

  • What is your goal? – clearly state what you want to know.
  • Context – give AI additional information so it can understand what you need
  • Source – describe the data where AI should look for a more focused response, and
  • Expectations – provide the specific conditions or requirements for the response.

The prompt can be as short or lengthy as you need it to be to encapsulate the four principles above – the more detail, the better. AI is quite capable of interpreting nuanced questions about complex business situations that include several sentences of qualifying information. The structure of the prompt is also important. For example, Microsoft suggests that Copilot may give more emphasis to the later part of the prompt compared to the earlier section (it is a teenager, after all).

The Detailed Prompt

AI will offer good results when you write, “Write a 1,000 word blog for an audience of publishing executives who want to understand potential conflicts and synergies between (1) the recent EU AI Act, which focuses on the safe technical development of AI, and (2) the EU’s own GDPR Act, which defines privacy protection rights for individuals and organizations. Note that much of the training data for Large Language Models may have been indiscriminately scraped off the internet without regard for privacy or copyright laws.” You will get less valuable information if you merely ask, “Tell me about the EU’s AI and GDPR Acts.”

If you do not get the response you are looking for at the first attempt, you can ask Copilot to revise or regenerate, and you will often get the answer you need. Think of it as a conversation which can progress based on prior answers. And again, before you use anything produced by any chatbot, make sure you verify the output against other mature and trusted sources.

Creative Prompts

The same four principles apply to more creative questions posed to AI, but in this example, you are asking for a different kind of response and where verification may not be quite as important. Here you might write a prompt such as, “You are an experienced marketing professional, and you need to generate creative ideas for a software company’s social media campaign aimed at medium-sized publishing companies with a readership of young adults focused on lifestyle and sustainability issues. The tone of the campaign should be informal and casual, aimed at publishers’ editorial and marketing personnel, utilizing the international publishing events calendar, and the SEO (Search Engine Optimization) keywords in the attached files.”

Here, we have asked AI to adopt a persona, in this case a marketing executive, who would understand the industry and the expectations of the targeted individuals. It is helpful to use ‘action” verbs such as “create,” “generate,” “imagine” or “draft” to drive actionable ideas. The format of the response may or may not be important, as the objective is to generate thought-provoking ideas. However, if you tell AI what format you want, it will comply.

In summary, tell AI clearly and specifically what your objective is, give it context and backup data, and if needed, tell it where to look for valuable information. Tell AI what perspective (persona) the answer should come from and to whom it is directed. And lastly, tell it how you want the output to look. If you put in this kind of work, you will be rewarded with the information or the inspiration you are looking for – and always check what AI is telling you with sources that you trust!

Be sure to read our white paper, “Unleashing the Power of AI in Publishing,” where we provide valuable insights into how AI is reshaping the publishing industry, offering practical examples of tools and strategies that are driving innovation and growth.

Image by DC Studio on Freepik.