Generative AI and Peanut Butter and Jelly Sandwiches: How to become a better AI prompter
Unless you’ve been under a rock for the last two years, you’ve heard of GenAI.
As the legal and technology worlds race to understand and harness GenAI, it’s easy to get swept up in the excitement and promises of what these tools can do. it is also easy to get discouraged by the seeming lack of useful instruction manuals on how to use them. In fact, many of us are having a hard time making it do anything at all.
One of the things that helped me bridge the gap came from my very first college class.
Learning how to make a peanut butter and jelly sandwich
Over 25 years ago, I was sitting in my first college class for my computer science degree. 25 bright-eyed students filed in and took their seats, nervously excited as we began our collegiate journey. In walks the professor with a pair of shopping bags from the local grocery store. He gets in front of the class and unloaded his belongings – a large tub of crunchy peanut butter, an equally large bottle of grape jelly, and 2 loafs of average sandwich bread. A paper towel was laid down and a small variety of utensils followed upon it. With our attention, he began:
Computers, as you’ll learn in this class, are excellent at following instructions. The role you will take as programmers will require you to give those instructions in the form of computer code. The instructions must be accurate, they must be precise, and they must be complete in a way that I imagine none of you are comfortable with.
In today’s class, you will write instructions to make a peanut butter and jelly sandwich. I trust you all know what that is. Your homework will be to complete those instructions. In our next class, we will examine your instructions and put them to the test.
With that and leaving the lunch kit exactly as he laid it out, he left the classroom.
A few scribbled down some notes. A few were off to the races writing as fast as they could. Most of us were confused but did the best we can.
The following class, we assembled again – this time, the professor already present and the environment for our assembly already in place. And, true to his word, the professor sought a volunteer to begin the experiment. The first candidate was asked to go to the lectern and read his instructions, one at a time and the professor would follow them exactly as they were written.
Starting with that student and each of us following in our turn – we each failed. Many of us failed due to taking assumptions that weren’t true and that we didn’t bother to validate. Some failed by way of implicitly doing multiple steps without writing them down.
We spent the entirety of the next 13 weeks beginning to learn how to provide instructions to computers. We learned to be equally accurate and precise in crafting these instructions.
We also learned to think about what we were starting from, where we wanted to go, and to build our instructions as the bridge from our start to intended destination.
So, we return to the here and now. Every major IT company is shouting about the power of GenAI from the rooftops. It was a rare vendor at ILTA that didn’t have something AI or AI-adjacent in their product line.
At the same time, the number of people who tell me they or the people in the firms they work for don’t understand what to do with these tools. People recognize that it can do everything but have a hard time identifying anything to use it for.
Ultimately, the use cases for the general use of GenAI must come from each of us individually. The way I use my technology platforms is different than each one of my colleagues at KK. The tasks that I’m attempting to do are also often slightly different than those on my team. All of this is true for you and your users as well. With different starting points and different destinations, it’s a losing battle to expect there to be a master list of prompts that everyone in the firm can use and save time and money.
Going back to my story, how can we use the exact same set of instructions when you keep your peanut butter in the pantry and mine is on the countertop?
Using GenAI tools both allows and requires you to think about your relationship with your data differently. It allows you to discover and process data in ways that really aren’t fundamentally new – but they can perform the instructions they are provided in an amazing speed.
The process of forming a prompt should be familiar to legal practitioners. Starting an argument from predicate and using logic and the rules of systems to build arguments that advance the state of a position towards the desired end goal is the same process that one uses to envision and build prompts. These skills and that process are very beneficial to use in prompting.
While you don’t need to know how to program to use GenAI tools, you are still required to give instructions to the tool. In fact, when we’re talking about just the basic chat facilities of GenAI tools, one of the major benefits is that the language used is English and not some esoteric computer language.
Find tasks that take a few minutes to a few hours. Then think about how to provide instructions to give to the GenAI tool:
Does the GenAI tool have all of the information that it requires to do your task for you?
- Intrinsic knowledge – This is what the GenAI tool is trained on. Depending on the tool, this could be a general source of information or a body of legal-specific knowledge.
- Trusted repository searching – Copilot can see everything you can see in your M365 tenant, for example.
- Web search
- Files attached to the prompt
With that source asserted, how would you explain the steps of the task to someone who has never done the task before? What do you expect the outcome to be? Describe the solution to inform the output.
Some examples might be processing data found in a document or webpage into a table or graph, drafting a response to a source document based on defined rules and the output expected to conform to provided examples, or recognizing either patterns or discrepancies in patterns in provided material.
The portions of our days where our work is driven by a set of rules, is a place where GenAI tools might prove helpful in helping us do that task more efficiently and possibly more effectively with a lower error rate.
Learning how to use these tools is an effort and is a skill that must be developed. And this is a process that many of us have undertaken before. 25 years ago, there was a new company that promised to facilitate the use of our data in new and unimagined ways. That company was Google and the product was the search engine. Those of us old enough to have gone through that transition remember staring at the empty Google search page with the same blank expression that I see many of my clients approach any of the GenAI tools when they first start using them.
Eventually, we learned how the product worked and most, if not all, of us are comfortable using it in our daily lives. In time, GenAI tools will take the same place as a piece of technology that we don’t remember living without.
As you think about how to write instructions to have GenAI tools do something meaningful for you, how would you build your peanut butter sandwich? Our professor’s entry clocked in at 64 pages of single-spaced print that he shared after going through our attempts.
That helped me have an appreciation for the level of detail that instructing computers require. I hope that by sharing this story, I can help you develop an appreciation for how to work better with GenAI tools.
To continue the conversation with the Kraft Kennedy team, please contact us.