Debunking the Myth of the Prompting Cheat Sheet // I Built an AI Brand Strategy App – Here's What I Learned (Part 1)

Throughout 2023, I embarked on a journey to build several AI applications, culminating in the creation of a brand strategy app. My primary motivation was to enhance my programming skills and explore the vast potential of the OpenAI API. This journey offered me deep insights into LLMs, particularly ChatGPT, its various models, and their functionalities (including their limitations).

I realized that many share the questions and misconceptions I had. In this series of posts, I aim to share key insights that I found particularly enlightening. Hopefully, you'll find them beneficial too.

Part 1: Debunking the Myth of the Prompting Cheat Sheet

Overview:

1. Model Selection: ChatGPT-3.5 vs ChatGPT-4
2. Function Calling: Integrating 3rd party data sources
3. "Temperature" Setting: Adjusting AI creativity
4. Setting "max tokens": Balancing data input and output
5. Utilizing my own database: Efficient data storage and usage

You might wonder, "Why bother creating a standalone brand strategy app when ChatGPT is readily available and free?" Well, it's crucial to understand that ChatGPT, as we use it on our devices, is merely an interface designed by OpenAI to interact with their sophisticated LLM. This interface, however, doesn't tap into the full capabilities of the AI, and simple or even complex prompts might not fully leverage the AI's potential.

Consider these aspects when using ChatGPT and the additional features the OpenAI API provides beyond the standard ChatGPT interface:

1. Model Selection: ChatGPT-3.5 vs ChatGPT-4

Whether you're aware of it or not, the choice of model is crucial. Using ChatGPT through a free account means you're accessing ChatGPT-3.5, suitable for basic tasks and gaining a preliminary understanding of ChatGPT. However, the paid GPT "Plus" plan grants access to ChatGPT-4, which boasts additional features like image interaction, voice-to-text, and image creation. More importantly, ChatGPT-4 generally outperforms ChatGPT-3.5 in delivering more accurate and sophisticated results. For instance, if you need consistently formatted results like:

$$$$$[b]{this is headline1 for <brandName>}[/b]$$$$$        

ChatGPT-4 handles such instructions seamlessly, unlike ChatGPT-3.5, which might struggle regardless of the prompt quality. In essence, ChatGPT-4 offers more consistent outputs, handles complex tasks better, includes more up-to-date training data, and manages larger text volumes. However, even with a paid subscription, ChatGPT-4 usage is limited to 40 messages every 3 hours. In contrast, the OpenAI API provides unlimited (paid) access to all models (because there are actually more models than just ChatGPT-3.5 and -4)

2. Function Calling: Integrating 3rd party data sources

For my app, I needed ChatGPT to analyze "Google Trends" data for specific keywords, timeframes, and regions. A standard prompt would merely suggest visiting the Google Trends website. However, with the OpenAI API, we can utilize "Function Calling" to directly integrate data from third-party APIs into ChatGPT. This feature allows for real-time data incorporation from various sources, such as Google Trends, stock market data, or recipes, limited only by the capabilities of the external API.

3. "Temperature" Setting: Adjusting AI creativity

The "temperature" setting is a powerful tool for adjusting the AI's creative output. By setting a score between 0 and 1, you control the randomness of the responses:

• A setting of 1 encourages more random and creative responses.
• A setting of 0 is ideal for factual tasks like data extraction, as used in my brand strategy app.

It's important to note that this adjustment cannot be made through prompting in ChatGPT; it requires access to the OpenAI API.

4. Setting "max tokens": Balancing data input and output

The "max tokens" setting is crucial for managing the balance between input and output data. For instance, a simple task like "write me a blog article" might require minimal input but generate a lengthy output. Conversely, a task like "summarize this blog article in one sentence" involves extensive input for a brief response.

Each GPT model has a token limit for input and output, typically split 50:50 by default. Adjusting this ratio can significantly enhance task efficiency, as I found with my app, which often required extensive input analysis for concise outputs.

5. Utilizing my own database: Efficient data storage and usage

While it may seem obvious, it's worth emphasizing: the standard ChatGPT interface is not designed for data storage. For further data manipulation, external tools like Zapier or manual copying and pasting are necessary. By integrating ChatGPT into my app, I gained complete control over data storage and management, enabling seamless task automation and ensuring consistent, efficient prompting.

If you have any questions or are interested in learning more about the UprisingBrands Strategy App, please feel free to contact me!

Coming Up:

• Part 2: Understanding the Limitations of GPTs
• Part 3: The Overrated Value of Proprietary Data
• Part 4: Preferring Indicative to Deterministic Results
• Part 5: Costs and limitations of GPT powered apps