“Let’s have a Chat(GPT)”— it’s about dialogue, not prompts

After all, it’s not called AnswerGPT

ChatGPT wasn’t designed as a simple answer generator but rather as a conversational partner, and that may be the best way to get high-quality results.

Since the introduction of ChatGPT, a new piece of jargon called “prompt engineering” has been popularized as the power-user method for working with Generative AI.

Prompt engineering involves crafting a precise and refined set of instructions that will reliably generate the best results from ChatGPT. A well-engineered prompt will include details about tone, format, writing style, and other considerations that guide ChatGPT in generating its output.

And it works. By asking ChatGPT complex, precisely written questions, you have a higher chance of receiving a response that meets your needs.

But it can be time-consuming and labor-intensive to draft and refine single prompts to achieve nuanced responses. Fortunately, there’s a simpler way: just start a conversation.

‘Few-shot’ vs. ‘one-shot’ prompting

ChatGPT was designed to engage in human-like conversations. And like a human, it prefers to have a conversation about a work assignment rather than just being handed a set of instructions.

Rather than providing a single “one-shot” prompt with all the instructions, testers have found that ChatGPT often performs better when given instructions that are divided into individual steps, through a practice known as “few-shot” prompting.

In few-shot prompting, you break up a task into separate logical pieces, similar to describing complex instructions in detail on paper, and pass them to ChatGPT through a back-and-forth dialogue. This approach allows ChatGPT some space to “think” about the content as it is being received, potentially improving the final response generation.

An example of few-shot prompting for fundraising might look like:

You: I want you to act as a helpful fundraising assistant. I will provide you with several sections of text that describe the background information about a fundraising campaign. Then I will ask you to generate several pieces of content using the information I have provided.

ChatGPT: Understood, please provide the first piece of content

You: Here is the first piece: This fundraising campaign is for Red Cross International. The mission of Red Cross International is .. <fill in mission statement>. In particular this campaign focuses on … <details about current campaign>

ChatGPT: Understood, please provide the next piece of content

You: Here is the next piece: The purpose of this fundraising campaign is to re-activate past donors whose gifts have lapsed. <more details about target audiences>

ChatGPT: Understood, please provide the next piece of content

You: That is the last piece of content. Now I want you to generate a suitable email fundraising appeal for this campaign. The message should be in the form of an email fundraising appeal. As much as possible, use the following guidelines for writing the message:
The email text should be approximately 350 words in length.
1) write an attention-grabbing opener at the start
2) Focus on the impact of the organization’s work
3) Include a compelling story
4) have a strong call-to-action
5) include a post-script that reinforces the main call-to-action.

Please list at bottom all source references for documents quoted or cited.

ChatGPT: Okay, here is the text of the email fundraising appeal:
<text follows>
…

In practice, this approach feels more like collaborating with a capable junior assistant than interacting with a computer.

The dialogue doesn’t have to end there. After ChatGPT generates the appeal copy, you can continue the discussion by providing feedback and suggestions for improvement based on your knowledge of successful fundraising strategies.

Through this iterative back-and-forth dialogue, ChatGPT can be guided toward generating the best version of fundraising copy.

Logical Frameworks

Few-shot prompting can follow different logical frameworks to hone in on specific outcomes, but the base idea of a back-and-forth dialogue remains the same.

1. Chain of Thought: This approach guides the model through a series of logically connected prompts to arrive at the desired output.
2. Incremental Detailing: Start with a high-level prompt and then add more details in subsequent prompts to describe complex tasks.
3. Socratic Method: This method uses a series of questions to guide the model toward the final answer, encouraging deep, critical thinking.
4. Example-led Prompting: Here, the model is given one or more examples of the desired output to demonstrate what is required.
5. Counterfactual Prompting: This approach poses hypothetical scenarios to the model and asks it to infer or predict outcomes.

These frameworks can be mixed to some degree within a dialogue. For instance, the example above combines elements of Chain of Thought and Incremental Detailing, but could also have included samples of previous fundraising appeals (Example-Led Prompting), or even Counter-factual Prompting, such as:

“Imagine if we had not received any donations last year for our charity organization. How might you describe the potential impact and write a persuasive appeal for donations based on this imagined scenario?”

Few-shot prompting is an evolving practice, and there may be more effective approaches in the future. For now, it’s a viable starting point for combining human fundraising expertise with the power of Generative AI.

It’s important to note that ChatGPT currently doesn’t retain working memory between conversations. ” So when you start up a new chat, you will be back to square one — ”You are a helpful fundraising assistant …”