Get More out of AI chatbots — 10 levels of prompts — The Business of Prompts
Created by Rob Tyrie CEO Ironstone Advisory Aug 8, 2023
CHATBOTS are not new. They are years old… supercharged smart ones are months old. The first implementations based on early Generative AI were a couple of years ago when some of the lingo that’s used with them were developed — like prompts and prompt engineering.
Why do they even call them prompts anyway? This is the vestige of the fact that the front-end interface to the large language models is based on natural language understanding and natural language processing. It’s in the spirit of a human prompting the large language model to generate a response with some sort of goal. If you heard the term “zero shot” or “one shot”, these are examples of simple prompt methods.
What follows are just examples of simple prompts of which there’s an expectation not of an answer from a database but a generation of information that imitates what a human would do with the information inside its language model. I suppose they could have called it the question or the data entry or the query even… But this new science has new terms. And in this case, it’s thoughtful we are just prompting something and expecting a conversation to occur. What enables the conversation is the fact that as you prompt a chatbot and it responds, it remembers your prompt and its response in the session.
Think of it as a kind of short-term memory so that you can build on the information in conversation with a computer. This is probably one of the more exciting aspects of this technology that is underrated. Human beings are not used to having conversations with computers, other than in science fiction novels and movies across the last five decades. In fact there used to be something called the Alexa prize, which was a prize given to a project that could hold a judged conversation with a human to entertain that human.
The longest conversation would be the winner. The conversation was just as entertaining by professional conversationalists like talk show hosts and actors. Back in 2017, I went to a presentation of the winning chatbots. This is where I realized that it would take multiple types of algorithms to achieve this as well as specialized kinds of data sets.. even back then the winner could have a conversation of 15 minutes of chit-chat with a person at breakfast as a scenario. We haven’t seen the chatbots connected to the series and the Alexis yet but we will. This is coming soon and we’ll use prompts with them as well.
In the research, there is detailed work going on to discover a better structure for interactions with chatbots so that the results are more accurate and error-free. If you’ve heard of terms like Chain of Thought, Tree of Thought or Skeleton of Thought approaches, these have been described in scientific papers and are formally defined in scientific language that tends to be obtuse. They are structured with a foreknowledge of how large language models and natural language understanding occur, so there is jargon within the descriptions.
I wanted to simplify approaches so that it’s easier to understand.
Start simply by adding some complexity to what you enter as a prompt in stage, you can narrow and focus the use of artificial intelligence to satify your needs.
If you’re really doing research it’s best to think of dealing with this form of artificial intelligence as having a conversation with the information. The conversation should illuminate the subject of which you are interested whether that’s creating something that is a synthesis of things or just explaining something that you wish to know more about. In all cases, these tools are best used in the hands of experts if they are to be used in a business context with material impacts.
It is not acceptable to blame machines for errors if you are a manager or business professional. You can use the tools just like you’d use the internet or a word processor but you’re signing the work so check it.
If you run the same prompt multiple times, you will find that different answers that are similar will be generated. Do not be surprised by this this is the nature of the large language model and how it is constructed to work. Small changes in your prompt like adding adverbs, or even punctuation will also alter the responses that are generated. In the spirit of checking your work, it’s okay to run the same prompt a few times and ask the AI to check its work, that is because now the response is in short-term memory for the session and it can be used as input.
10 Levels of Creating a prompts
Here are 10 levels of increasing complexity for prompt engineering, with examples, for research assistants and business managers:
Level 1 - Keywords
Use a few key nouns to indicate a very broad topic, like "marketing strategy" or "microbiology research". Gives the AI wide latitude to talk generally about the subject.
Level 2 - Buzzwords
Include some key buzzwords and jargon to point the AI towards a more specific aspect of a field, like "LEAN manufacturing" or "CRISPR gene editing". “Give me the business canvas for Airbnb. This narrows things down slightly and gives a target form for the results.
Level 3 - Single Sentence
Craft one sentence clearly and literally summarizing the task or objective, such as "Summarize the key benefits of implementing a just-in-time inventory system" or "Provide an overview of biofuel production from algae." Further information constrains the scope but leaves room for the AI to complete an interpretation.
Level 4 - Ask a Question or 2
Pose a focused research question or problem statement, like "What is the current scientific consensus on factors affecting enzyme production rates in bacteria?" You can use critical thinking here and use the 5 W’s. Or the 5 Why’s. This defines a specific issue to address and your need for an answer. Don’t forget the question mark.
Level 5 - User Story
Outline a typical user story in 1-2 sentences to provide business context, such as "As a product manager, I want to understand the pros and cons of different pricing models for our smartphone app. Provide viewpoints from the CFO and the CRO". Specifies a perspective for the AI to adopt.
Level 6 - Goal Statement
Clearly state the end goal or purpose in 1-2 sentences, like "Provide actionable marketing strategies for improving customer retention rates in the next quarter" or "Suggest experiments to test how temperature impacts chemical reaction kinetics." Defines task objectives and metrics for success.
Level 7 - Bullet Points
Provide 3-5 bullet points outlining background context, previous research, hypotheses, etc. to frame the prompt and reduce misinterpretation. Here is a set of points about doing competitive intelligence.
1. Identify your competitive segment (eg Marketing Automation)
2. Identify your key competitors within the industry.
3. Gather data ethically and legally using various open-source methods.
4. Analyze data to extract valuable insights and trends.
5. Formulate actionable recommendations to enhance competitive advantage.
Level 8 - Paragraph Prompt
Construct 1-2 paragraph prompts explaining scenarios, assumptions, source materials, and methodology in detail. Example
“Create an engaging and impactful social media campaign plan for a new eco-friendly lifestyle brand. The brand focuses on promoting sustainable products and practices to inspire conscious consumerism. The campaign should target environmentally conscious millennials and Gen Z individuals, leveraging platforms like Instagram, TikTok, and Twitter. The goal is to raise awareness about the brand’s mission, drive organic traffic to the website, and foster an online community of eco-warriors. Be sure to incorporate eye-catching visuals, compelling storytelling, and interactive elements to encourage user participation and sharing. Additionally, include strategies for influencer collaborations, hashtag campaigns, and metrics for measuring the campaign’s success. The deadline for the plan submission is two weeks from today.”
Significantly constrains the AI's problem space.
Level 9 - Guided Template
Create a partially completed document, analysis, or template for the AI to finish filling out. Provides substantial structure and guidelines.
Introduction: This document provides an overview and analysis of the [Project Name], a software development initiative undertaken by [Company/Organization Name]. The purpose of this review is to assess the project’s progress, identify key achievements, challenges, and potential areas for improvement. The review encompasses the project’s objectives, scope, methodology, team performance, and adherence to timelines. The insights gathered from this evaluation will serve as valuable feedback for future projects and help optimize the overall software development process.
Project Overview:
Project Name: [Project Name]
Start Date: [Date]
End Date: [Date]
Project Manager: [Name]
Team Members: [List of Team Members and their Roles]
Stakeholders: [List of Key Stakeholders]
Level 10 - Completion and Editing
Give the AI a nearly complete work product and have it generate minor additions like a conclusion, executive summary, citations, or data visualizations. If the document is long you may have to use a Chatbot with a larger prompt space like Claude.AI. This means taking your whole report, presentation or document and adding it to the session. Minimizes scope for misunderstanding.
The levels progress from broad topics to providing full frameworks and context. Higher levels reduce AI creativity but enable prompting for sophisticated, tailored outputs.
Good luck with your explorations and generations.
