1. GPT inputs
Great work so far. We've spent time learning the fundamentals of safe GPT use. It's now time to focus on how to get what we want out of it!
2. Prompt engineering
The text we input into a GPT is known as a prompt. Editing and tweaking this prompt to get what we want from the tool is known as prompt engineering. If our prompts are bad, the result from the GPT will also be bad.
3. Prompt interpretation
A GPT tool interprets a prompt by identifying the high-level topic and attempting to understand the request. For example, with the prompt "Write a product description for our new shirt",
4. Prompt interpretation
the tool will identify the high-level topic of a new product description,
5. Prompt interpretation
and the request is to write this description. While this is a good start, this prompt likely won't get us what we want.
6. Techniques
A few techniques help generate better results. These can be boiled down to being specific by including additional context or instructions,
7. Techniques
providing an example,
8. Techniques
and iterating over the results.
9. Be specific
Let's use these techniques to improve our original prompt. The GPT tool we are using likely has access to additional information about the product which the business has approved for the tool, or we may use image software with a built-in GPT tool.
10. Be specific
This allows us to be more specific by including a product or file name.
11. Be specific
We may have specific things in mind that we want to highlight, like a new color or design, so we include that context in the prompt. We might also need it to fit a specific format, so we request that the description be written in bullet points
12. Be specific
and include a summary sentence.
13. Provide an example
If there is a product description we particularly like, we can use it as an example in the prompt.
14. Provide a template
We can take this further by providing a template. Some tools allow a template to be attached for automatic completion, while others may require us to upload one. Throughout the process, we must always consider the earlier mentioned privacy concerns.
15. Iterate
Lastly, we can iterate over the result by going back and forth with the GPT with some additional tweaks, such as using a specific phrase, adjusting the length, or changing the tone.
16. Confidential tasks
But what if the task involves confidential information, such as using a GPT tool within a legal workflow? While the tool will have built-in safeguards, given the sensitive nature of the request, it is recommended that legal expertise is incorporated into the process so that prompts are vetted and legally sound before they are put into the GPT tool.
17. Effective prompts
To sum up, if we follow these practices of being specific, providing context, instructions, or examples, and iterating over the results, with the added safety of vetting prompts that handle sensitive data, then we'll end up with an effective prompt.
18. Let's practice!
The good news is that we get better at this the more we practice it, so let’s do that now!