The art of prompting AI: Science or Witchcraft?
While there are no secret prompts that will unlock new levels of responses, you will get vastly different results depending on how you prompt and which language model you use
Paul Cheung is a RJI Columnist who will explore how newsrooms and journalists with limited resources can use AI tools to improve efficiency and increase capacity in their daily development and audience operations.
For small and remote news organizations like The Center for Public Integrity, staff often wear multiple hats and don’t always have a brainstorming partner available to them at all times. So, I now start with ChatGPT to help initiate brainstorming new ideas or help me explain a 5,000 word investigation in a paragraph to donors and funders without asking our editors to customize 1,000 different versions.
To unlock the power of these predictive language models, I researched best practices. You will be bombarded by a lot of self-proclaimed experts when doing this. In fact, “prompt influencers” are popping up like wildflowers across all the social platforms, and they will provide expert advice on almost anything under the sun, from how to use ChatGPT to develop a marketing plan to starting a business plan for your next side hustle to even writing a book!
So that begs the questions:
- Is there an art to prompting?
- Is it necessary?
- Does it matter to journalism?
The short answer to all of these questions is conditional yeses.
First, what is prompting? Prompting is the process of providing an initial message or context to a language model like ChatGPT or Bard to guide its response.
While there are no secret prompts that will unlock new levels of responses, you will get vastly different results depending on how you prompt and which language model you use. ChatGPT and Bard will give you different output despite using the same prompts.
Here is an example of three prompts I used to summarize a CPI story about how forcing more people to work for food stamps is harmful.
Prompt 1
Act as an editor for a major news publication and summarize this article in one short paragraph
- Basic Formula: Role (Editor) + Action (summarize this article) + Output (one short paragraph
- Result:
Title: Study Finds Imposing Work Requirements for Food Stamps Harms Vulnerable Populations
According to a recent study, the implementation of work requirements for individuals receiving food stamps has proven to be detrimental. The report reveals that such policies lead to adverse consequences, including increased food insecurity, diminished access to healthcare, and worsened overall well-being for low-income Americans. The findings shed light on the need for a reconsideration of these requirements, emphasizing the importance of alternative approaches to support individuals and families in need. - Observation: The summarization is general and somewhat vague.
Prompt 2
Act as an editor for a major news publication and summarize this article for high school students in one short paragraph
- Formula: Role (Editor) + Action (summarize this article) + Target (High-school students) + Output (one short paragraph)
- Result:
Title: Study Shows Negative Effects of Work Requirements for Food Stamps
A recent study discovered that requiring work for individuals receiving food stamps can have harmful consequences. The research found that these policies lead to increased difficulties in finding enough food, reduced access to healthcare, and overall negative impacts on the well-being of low-income Americans. This study highlights the importance of rethinking these requirements and finding better ways to support those in need. - Observation: Same length as prompt one except the vocabulary used was much simpler because of my specification of the audience.
Prompt 3
Act as an editor for a major news publication and highlight three key takeaways of this article for a group of McKinsey consultants
- Formula: Role (Editor) + Action (highlight 3 key takeaways) + Target (McKinsey Consultants). Note that I’m being very specific about what I want here.
- Result: Title: Key Takeaways: Impacts of Imposing Work Requirements for Food Stamps
- Adverse Consequences: A comprehensive study on the effects of work requirements for food stamps reveals alarming outcomes. Imposing such policies results in increased food insecurity, limited access to healthcare, and overall worsened well-being for low-income individuals and families.
- Need for Reconsideration: The findings strongly emphasize the need for policymakers and consultants to reevaluate the effectiveness of work requirements. Alternative approaches should be explored to better support vulnerable populations, ensuring their access to basic necessities and improving their overall quality of life.
- Holistic Support Systems: The study underscores the importance of comprehensive support systems that address the root causes of poverty and unemployment. It highlights the necessity of combining access to food assistance with initiatives focused on job training, education, and healthcare, in order to uplift individuals and families from poverty more effectively.
By considering these key takeaways, McKinsey consultants can contribute to shaping policies and strategies that prioritize the well-being and long-term success of disadvantaged communities.
- Observation: While not short, it got a lot more specific and the details are far more nuanced than the previous two versions. Also notice how ChatGPT inserted their own “opinion” of what people should do in the output.
These examples illustrate that there is some art and science to prompting. A successful prompt follows a basic structure comprising a combination of role, action, target audience, and format. To achieve optimal results, it is recommended to keep prompts concise, specific, and straightforward, akin to explaining something to a 5-year-old because they don’t necessarily understand the full-context of what you want. I would also treat any outputs/responses from these language models like a suggestion from a very casual acquaintance and not a best friend; extract the part that you deem useful and disregard the rest.
This is all very temporary for two key reasons
First, these language models are evolving rapidly and are expected to better understand intent without relying heavily on precise prompts.
Second, while the marketplace is still very nascent with ChatGPT and Bard as dominant players, it won’t stay that way for long. More competitors are expected to come, according to this VentureBeat article.
So for now, becoming a prompt expert is not a necessity for journalists, so you don’t have to waste precious resources training your entire staff on prompting or hiring prompt consultants. But it is nice to have the knowledge at the moment if you want to use it.
For those who are interested in learning more, here are two resources you can explore relating to prompting:
- https://www.promptvibes.com/: You can sign up for a free newsletter and search prompts by categories from the homepage.
- https://www.promptstacks.com/: They offer vetted prompts as well as a Discord community group of prompt engineers. They also offer virtual courses like ChatGPT Course: The Comprehensive Beginners Guide for $24.95.
Comments