GPT “Two-Chat” Method

Optimizing ChatGPT Prompts with Allie McCarron and Emily Witucki ??

Navigating the world of AI and prompt engineering can be complex, but with the right strategy, it becomes a breeze. Countless courses exist to help you refine your interactions with large language models (LLMs) in AI, but sometimes getting started causes a state of overwhelm: how do I address the ‘bot? What’s the right question to ask? How can I formulate my ask in such a way that it conveys the correct level of formality and necessary formatting in the output?

The above encompassed only a few of the many questions asked by fellow learners in our recent Clicked AI Quest, where we focused on incorporating these artificial models into our routine of solving problems in Salesforce for everyday clients.?

So how to get started? Here's a simple yet powerful method we used to ensure our prompts to ChatGPT generate the best possible results for ourQuest.

Step 1: Drafting the Initial Prompt

• Open a chat window with ChatGPT or your preferred LLM (Bard, Claude, et cetera).?
• Begin by asking ChatGPT for help explicitly. Within your initial request, clearly state that you need it to write a prompt for ChatGPT. This specific instruction is crucial, as it directs the AI to focus on creating an optimized prompt tailored to its own processing capabilities. In addition, you’re leveraging the AI's understanding of its own processing and optimization techniques to receive a more effective prompt tailored to generate your desired outcome(s).
• Next, articulate your request or question with as much clarity and specificity as you can manage. Instead of “I need assistance with incorporating APEX code into a flow,” try “I need some APEX code for Salesforce that can do the following: query all standard objects within the org for Developer Name, API name, and number of records associated with that object. In addition, I will need to incorporate this into a screen flow.”?

In our applied example for Clicked, we used a blend of five different parameters to help us generate a clear and specific initial request for ChatGPT:

• Role (what persona will the AI emulate?)
• Request (what do you want the AI to do?)
• Context (what are the circumstances for the request?)
• Specifics (what details do you want to be included or excluded from the output?)
• Format (how do you want the output organized?)

By setting up this initial request properly, the below output provided a basis for a more refined (as well as complex) query for Step 2.

Step 2: Applying the Refined Prompt

• Open a new chat window with ChatGPT. Don’t skip this step, as it ensures that the context from your initial request won’t influence the new interaction.
• Enter the refined, optimized prompt you received from the first chat. Now crafted with the AI's insights, this version will yield responses or solutions within a narrower field of view.

So, what happened as a result of this method???

When we used a refined prompt asking ChatGPT to analyze a stakeholder interview transcript, our output became more structured and insightful compared to a general 'analyze this transcript' request. For example, we received a concise summary of the client’s needs, followed by more detailed requirements for each primary topic (marketing, analytics, and CSAT scoring).

In contrast to the vague and scattered responses typical of generic requests (such as “based on this interview transcript, what does the client want?”), this strategy provided a comprehensive analysis that included exactly the specified criteria in Prompt 1:

• Summary of project scope
• List of specific stakeholder needs
• Pain points in the current instance
• Essential KPIs

Not only did we receive data output specific to our requirements, but ChatGPT also yielded distinctly actionable insights from the transcript in a relatively short amount of time compared to someone manually creating notes from a transcript or recording. Granted, we spent some additional time checking over the AI’s work and refining it, but successful engineering via the Two-Prompt Method saved our team several precious hours that could then be spent on other valuable tasks.

Don't be afraid to treat ChatGPT as if you were an English teacher – AI is often prone to hallucinations or retrieving insights that don't match the input. Follow up with questions such as –

• What evidence do you have to support this section of your response?
• Show me in the source text where you found this?
• How did you arrive at this particular insight?

It's a machine, and it's constantly learning! Remember that the output is only as good as the input and binding parameters, so AI needs that human element to not only check its work, but also to encourage it to assume accountability for its own.

In addition, if the data output displays significant errors or inaccuracies, tell it so and force a re-computation. In our earlier example for creating APEX, if you plug the generated code into Salesforce and find that it doesn't compile correctly, enter the error directly into ChatGPT if you can't locate the answer yourself.

Why This Works: This two-chat system effectively utilizes ChatGPT's own 'expertise' in understanding and responding to prompts. It's a feedback loop where the AI assists in formulating a question that it can answer more effectively, ensuring clarity and relevance in the responses.

Benefits of the Method:

• Enhanced Accuracy: By refining prompts with the AI's input, you're more likely to receive precise and relevant responses.
• Efficiency: Saves time and reduces the trial-and-error process of prompt crafting.
• Learning Opportunity: Observing how ChatGPT rephrases your initial prompts can provide insights into effective prompt engineering strategies.

What other methods or strategies do you use for prompt engineering and output generation with large language models such as ChatGPT?

Until next time, live long and automate.