Why the “Prompt” Is the Key to Unlocking AI Success

Imagine you have a chance to meet someone you greatly admire, whether it’s a business leader like Marc Benioff, or a celebrity like Taylor Swift. The key to making the most of this moment is to ask them insightful questions. I recall my own experience learning bass guitar from a remarkable teacher, a member of an all-female Led Zeppelin tribute band and a renowned session musician. I told her: “I don’t just want to improve my bass skills.” She was initially surprised, but then I explained: “I want to learn how to think like a musician, the way you do.” This changed her teaching approach, and accelerated my learning too.

This idea is also crucial when using Generative Pre-trained Transformers (GPT). The true strength of GPT doesn’t come from its intricate algorithms or its Large Language Models. It’s all about the prompts you use and the way you frame your questions.

Smart Prompting: Crafting the Right Questions

Think of smart prompting as the insightful questions you would pose to someone you deeply admire. With Large Language Models like ChatGPT, Claude, and Bard, it is the quality of your questions that is crucial, and the additional context that you give it as part of the prompt.

GPT is like having a zealous, quick-learning intern at your service. For example, if you asked the intern to “book a restaurant”, hopefully they will come back with further questions – what, when, why, who, where – rather than just booking it. If the answers are equally vague, there are yet more questions, back and forth. But, if you specify, “Find a restaurant for a business lunch with a client who enjoys Indian cuisine, where we can discuss privately. He’ll be in town on Tuesday the 3rd,” you’ll probably get the right answer the first time.

The same is true for GPT. The key lies in the specificity and length of your prompts. The more detailed and longer the prompt, the more comprehensive and accurate the response. However, the challenge lies in consistently inputting detailed prompts of 3-4 paragraphs or more. The prompts could be longer than the answer. This may not be worth it for a one-off, but if the prompt is going to be used multiple times or by lots of people, then the prompt is worth refining over time to improve the outcomes.

Our Elements.cloud Product Management team, for instance, uses prompts that are 300-400 words long, refined over the last 6 months. These detailed prompts have significantly enhanced our team’s productivity and we are seeing 20-50x gains. They’ve become an integral part of our daily operations. You can also see that the prompt has text that is between the {} that is replaced before the prompt is used. So it is really a prompt template. 

Prompt Templates: Operationalizing Use of GPT

The team has operationalized these prompt templates. They serve as a base, where specific sections are customized to meet unique needs before being submitted to GPT. So the team has been building a prompt template library.

Centralized storage of prompt templates is key. This allows team members to share and effectively utilize GPT across our organization, enhancing productivity. The development and refinement of these templates represent a significant investment, highlighting their value as productivity tools. Therefore, we need to manage the prompt templates just like code or metadata configurations.

Access to edit prompt templates should be controlled; this prevents the loss of refined content through accidental alterations, similar to restricting admin rights in Salesforce. Additionally, providing feedback to the template editors is crucial for ongoing improvements, necessitating version control. It’s also important to track the usage of these templates, akin to monitoring a Salesforce configuration.

Standard GPT, Custom GPT, and Apps Using API

There is a difference between the free ChatGPT service, ChatGPT Plus which has a $20/month cost, and the way that apps are accessing GPT. They are all using the LLM but driving it in slightly different ways. They are also enabling users and ISV apps to make the prompting more specific so that the answers are more accurate and have fewer hallucinations.

Standard GPT

The diagram below shows the standard GPT architecture. The User enters a Prompt in the ChatGPT UI. The LLM returns a Result. The only data it can draw on is the LLM, or the single file that you attach with the prompt. Hence it is subject to all the biases of its training data.

Note: These diagrams were drawn in the Salesforce Diagrams Architecture Standard using Elements.cloud.

Custom GPT

OpenAI has launched Custom GPTs that are tailored AI models that can be enriched with instructions, specifically up to 10 knowledge files and API links to external information. This means that much of the prompt that you used to need to enter could be saved as part of those instructions. It also means that you can give the model your specific information in conjunction with the LLM, giving you far better results. This is shown in the diagram below.

The UI is the same as standard ChatGPT. As the custom models use the instructions and knowledge in addition to the prompt it means the prompt can be shorter. Access to these models requires a ChatGPT Plus account which is $20/month.

ChatGPT Plus users can develop these Custom GPTs, choosing to keep them private or make them public. If they are marked private, they are accessible only through a unique link, ideal for sharing confidential or niche information with a limited audience, like colleagues. Public models are open for broader user access, serving as a platform for content creators to distribute information and tools to a wider audience.

The quality of these custom GPTs varies, influenced by the creator’s expertise, instruction quality, attached knowledge, and the effort taken to train and refine them. The most effective models typically have a narrow focused scope.

Examples include:

• TDX Submission Guide (Public): Coaches you on how to present your idea for a presentation when completing the submission form.
• Success Story Writer (Private): Takes a transcript of a customer interview and writes a success story in a specific format.

The effectiveness for the users of these Custom GPTs largely depends on the quality of the prompts. The creator probably knows the best questions to ask. A Custom GPT allows for four short initial prompts, but achieving great results usually requires far more detailed prompts.

Security Warning: If you make your custom GPT model public, it is possible to trick the GPT into revealing the instructions, listing the names of the knowledge files, and summarizing the information in those files. So think carefully about what data you are happy to have in the public domain, and the naming of your knowledge files before you make it public. Also remember that anyone with the private link can access that model.

GPT Embedded in an App

Salesforce and a number of ISVs are now embedding GPT into their apps. This means that the app is helping format the prompt and is tailoring the instructions and data it is providing to the GPT based on the prompt. The result may be presented back in the UI, or may be used inside the app. This is controlled through the app’s security model and UI.

Examples include:

• ElementsGPT can take the transcript of a discovery call, a screenshot of a process diagram from any app, or a hand-drawn process, and it will then design a process map. It is not just drawing a diagram, but it is using its knowledge of processes to design the diagram. 
• Metazoa and ElementsGPT both allow you to ask questions about your org data through prompts and get text results back.

Writing Great Prompts and Prompt Templates

Here are 10 principles for writing great prompts and prompt templates:

1. Prompt templates, not prompts. Aim to refine the prompt so that it is a reusable template. To do this you need to get an acceptable result from your initial prompt. It shouldn’t result in a back and forth conversation. If it does, then track the conversation so that you can incorporate the changes into the prompt template.
2. ChatGPT is not Google Search on steroids. ChatGPT is far more than a search engine. It’s a conversational AI designed for interactive, natural language discussion. To craft better prompts, approach ChatGPT as a partner in conversation rather than a search query tool. It’s about engaging in a dialogue where you build on responses, clarify points, and explore topics in-depth.
3. It’s great at solving puzzles. ChatGPT excels at tackling complex problems and breaking them down into manageable pieces. Code Interpreter is a ChatGPT function that looks at the prompt and the data and then writes code to do the analysis. When writing prompts, present your problem like a puzzle that needs solving. This means providing all the relevant pieces – details, constraints, and desired outcomes for example.
4. Set the context. Context is crucial. If you’re a business analyst, frame your prompt with that perspective. For instance: “As a business analyst, I need to understand the implications of X on Y. Can you outline the potential impacts and suggest analysis methods?” This helps ChatGPT tailor its responses to fit your professional scope and needs.
5. Be specific about what you are asking for. Vagueness leads to generic responses. Be as specific as possible about what you need. Instead of saying, “I need help with data analysis,” specify the kind of data, the analysis goal, and any particular methods or tools you’re considering. This precision guides ChatGPT in generating more targeted and useful responses.
6. How do you want the results delivered? ChatGPT can deliver the result as text, code, images, and now with Code Interpreter, it will draw charts. Specify not just what you want to achieve, but also how you’d like the results presented.
7. Set the standards. If there are industry standards, it makes it easier for the analysis of the input data and the formatting of the result. User stories have an accepted format… so do UPN process diagrams. Reference specific frameworks, methodologies, or industry standards in your prompts to align ChatGPT’s responses with professional norms.
8. Store and refine. Keep track of the prompts you use and how effective they are. Store successful prompts and constantly refine. This process of iteration will help you hone your skill in writing effective prompts. 
9. Operationalize – share, refine, track. Consider how your prompts can be made part of your team or organization’s operational processes. Share successful prompts, refine them based on collective feedback
10. Experiment. Finally, don’t be afraid to experiment. The field of AI is rapidly evolving, and what didn’t work yesterday might work today. Try different approaches, play with the phrasing of your prompts, and observe how slight changes can lead to different types of responses.

Managing Prompt Templates

You need a structured way of managing prompt templates that are used in ChatGPT and other models. Management of prompts for your organization could be replicated as a couple of custom objects inside Salesforce. It’s important to note that this is not a substitute for Einstein Prompt Studio, which is specifically for managing prompt templates inside Salesforce. 

As everybody is still learning, we want to encourage people to create and share Custom GPTs, but also share the prompts that make them work. But also to start storing their private prompts as templates. So we created a free app called GeePeeTee.cloud. It began as a directory for Salesforce-specific GPTs, designed to help the ecosystem easily locate them. However, we saw the need for prompt management so its functionality has expanded significantly. 

Summary

The real strength of GPT comes from knowing how to prompt it effectively. Using the right prompt is key, but do not underestimate the time to structure and refine prompts as templates that can be reused by your teams. Prompts as templates are valuable assets that should be managed like any other application. Custom GPTs, which are focused helper apps, enable huge productivity gains when teamed with customized prompt templates. Prompt management is definitely a new skill to be mastered.