When starting to prompt for large spreadsheet/data entry projects, the first few prompts may very well fail! Do not be discouraged - this is completely normal and part of the prompting process, as ChatGPT4 and plugin abilities for these projects vary widely based on multiple factors (e.g. how a company has data arranged on its website).

Based on the results of those first prompts, we learn how to best refine our prompting strategy for a specific project. Here you’ll find best practices for doing so, and how to implement those practices across several types of use cases.

Best Practices At a Glance 👀

• Use the GPT-4 with DALL-E, browsing, and analysis, or determine the best suited plugins based on the initial results.
  • If a plugin fails to perform the assigned task, try alternative plugins that perform the same task (e.g. AskYourPDF and ChatWithPDF, LinkReader and WebPilot, etc.)
• Determine the best structuring of instructions - whole research vs broken down vs excessively broken down?
• Provide the research strategy yourself.
  • For example, “use ChatWithPDF to find …. Datapoints in this report (link). If you are unable to find the data points in the pdf file, use Voxscript to search for the data on Google and use multiple search strings (could even provide search strings to look for) until you are absolutely sure that the data does not exist in the public domain.”
    • See the example here.
• If the model struggles with finding links itself and extrapolating data, provide the links directly.
  • See the example here.
• When plugins or the model can not pull data from a site, but you are able to copy-paste from the site (e.g. for a list of items or locations), copy-pasting from the site into the model and asking it to compile the information into a table is a big time saver, as seen in this example.

If these practices still fail, take a do-it-yourself approach, and take advantage of the model as an assistant whenever a new potential use case arises (e.g. write-up, new quick data searches, etc.) We’ll explore more below!

Best Practices: Deeper Dive 🏊🏼

Case 1: ChatGPT Plus struggles to find and provide the required findings

1. Conduct Preliminary Research: Before engaging with ChatGPT Plus, conduct your own preliminary research. By providing self-sourced references or guides, you can make the task more accurate and efficient, while saving the number of prompts available to you (i.e. 40 prompts in 3 hours for GPT-4).
2. Determining Suitable Plugins: Based on the results of the initial interactions, you will be able to identify the most appropriate plugins and determine if you need to change plugins for the project. Remember, different plugins have varying efficiencies for different tasks.
   • If a plugin doesn't successfully complete a task, try alternate ones offering the same functionalities. For example, if "AskYourPDF" isn't satisfactory, consider switching to "ChatWithPDF", or from "LinkReader" to "WebPilot".
3. Prompts and Instructions Structuring: Find the optimal structure for your instructions - should they cover the entire research project, be broken down into parts, or be excessively detailed? This may require some trial and error.
4. Provide a Research Strategy to Follow: If ChatGPT Plus is not able to use a good research strategy on its own, feed the research strategy into the model yourself.
   • Here is an example prompt showing how we can feed the strategy: “use ChatWithPDF to find …. (define data points) from this report (provide link). If you are unable to find the data points in the pdf file, use Voxscript to search for the data on Google and use multiple search strings (could even provide search strings to look for) until you are absolutely sure that the data does not exist in the public domain.”
5. Copy-Pasting Information into ChatGPT Plus: When plugins or the model can not pull data from a site, but you are able to copy-paste from the site (e.g. for a list of items or locations), copy-pasting from the site into the model and asking it to compile the information into a table is a big time saver. The model will sort out the required (important) information from the whole clutter and compile it into the required rows and columns that you can directly copy into a spreadsheet. See example here.

Again: If ChatGPT Plus is still unable to find or provide the required data, consider performing the full research yourself and use ChatGPT Plus as an assistant when new potential use cases, such as write-ups, quick data searches, compiling data into a table, data visualizations, or others, arise.

Case 2: ChatGPT Plus Starts Providing Relevant Data

1. Relevant but Unspecific Data: If the model begins finding relevant data, but not the exact data that you need, try breaking down the instructions into more detailed steps to guide the model better. See if the model follows your instructions properly.
2. Successful for One Item: If the model successfully finds the required data for one item/company/country, etc., the next prompt could simply instruct the model to replicate the same steps for the next item. This eliminates the need for re-providing or rewriting the whole prompt set and saves time and the prompting limit.
   • Here is an example of a follow-up prompt: “Follow the exact same steps as above and conduct the same research utilizing the same strategy for xyz (i.e. next item).”