The key to working effectively with AI is to consider the behaviors you are looking for, and write a prompt that will instruct the tool to give you those results.
A prompt is the instruction you give to an AI in order to elicit a particular response. For a tool like Google Gemini, this is the text that you would type into the prompt box when you start an interaction with the LLM. A prompt can be anything from a single sentence to an extremely detailed description of the output you want the tool to produce. Often, the more thought you put into crafting your prompt, the better results you will see.
Table of contents
Video Guide
Choosing a Prompt Structure
Working With a Prompt Structure
Iteration
Custom Bots
Prompt Libraries
Video Guide
Choosing a Prompt Structure
It can be effective to work with a prompt structure that cues you to fill in different components to initiate that conversation, and several effective structures have been developed and published online. Some examples are the CLEAR and COSTAR frameworks summarized in this guide from the United States Military Academy Library. We will work below with another simple prompting framework, but note that many interesting options exist, and which works best for you might depend on your goals and preferences.
More involved prompting strategies can yield good results for complex or specialized tasks. When addressing a complex question, an interesting approach can be to explicitly direct a chatbot to take multiple angles on the problem, for example using Edward de Bono’s ‘six thinking hats’; this can yield results that are more multifaceted and nuanced than a simple prompt. Prompts can also be highly engineered for specialist work: an example of a prompt that very deeply specifies the work to be done is the Sift Toolbox created by Mike Caulfield, which instructs an LLM to behave as a sophisticated tool for fact-checking and contextualization, and write structured reports on its findings.
Working with a Prompt Structure
The below example is based on a model quoted in the Student Guide to Artificial Intelligence from Elon University and AAC&U, which was originally developed by AI engineer Ben Hylak:
- Your goal
- Task instructions and form of output
- Constraints, warnings, and verification requirements (for example, referencing)
- Background and context – upload or link to specific resources to work from
- Optional: The role the AI should play
- Optional: Examples for the AI to follow
This prompt structure asks a user to establish which materials an AI should work with, specify the goal of the conversation and how the AI should behave, and consider what will be needed to ensure a reliable result.
An example prompt from the above guide is this:
“I am presenting to a group of college students about ethical issues related to artificial intelligence. I need help identifying and organizing ideas.”
“Identify the most important ethical issues related to AI. Create a list of 8-10 key points to make, with several bullets for each point explaining its importance. Cross-reference your recommendations with classical principles of ethics and draw connections between those principles and emerging AI technologies. Suggest visuals for the presentation.”
“Use the points in my attached paper. Provide links to all sources. Draw material from respected academic publications only.”
“For context: I have been studying AI ethics and have written the attached paper. I now need to explain these issues in a 15-minute oral presentation. The students have only general knowledge about AI and most have not studied ethical matters.”
“In providing your list, consider yourself to be the presenter, who is a college-aged student very close in age to the audience.”
“I have created a start on the presentation and have attached that to this prompt. Please provide a critique of my ideas so far.”
In a learning context, it is worth noting that generating finished materials is only one purpose to which an AI can be turned. More interactive uses may be much more beneficial for learning. When asked a direct question about a topic, an AI will often respond by generating paragraphs of text that answer the question, where a human educator might proceed by asking questions to help a student to come to the answer on their own. An AI can, however, be prompted to proceed in a much more interactive way that might benefit learning.
Here is an example prompt adapted to promote a more interactive usage:
“I am presenting to a group of college students about ethical issues related to artificial intelligence. I need help developing my understanding of the material and practicing for the Q&A.”
“Ask me five difficult questions about the material, one at a time. Once I have answered each one, ask me further questions to help me to see weaknesses in my response. Then, give me feedback: let me know which points I covered well, and what I could say more about.”
“Use the points in my attached paper and slides. Provide links to all sources. Draw material from respected academic publications only.”
“For context: I have been studying AI ethics and have written the attached paper. My audience will be knowledgeable about argumentation around AI, and may have advanced knowledge of neural networks, which I find tricky to understand.”
“Write your questions as though you are a computer science student.”
“Use these example questions to get an idea of the level of complexity needed.”
For more detailed exploration of possible applications, see Applications for Student Learning, Assignments in the Context of AI and Labor-Saving for Educators.
Custom Bots
Google Gemini offers the ability to create your own ‘Gems’: custom-built, user-created chatbots created by providing specific source materials and instructions. If there are prompt elements that you foresee yourself using regularly, such as a particular voice or persona and a set of guidelines for creating or checking outputs, you can load those into the Gem’s instructions, and create a custom chatbot cued into those behaviors that you can come back to at any time.
For more instructions on setting up your own custom Gems, visit the page on Google Gemini/Gems.
Iteration
It is very important to go in with an expectation that you will need to iterate. If you do not get useful results the first time, consider why. Adjust your prompt, try again, and observe the results. An AI will often not be employing all of the contextual assumptions being used by its human interlocutor. A human questioner will often find that their original prompt takes an AI in a slightly different direction than they were hoping, and the process of refining and adjusting prompts can itself clarify a person’s understanding of the material.
Prompt Libraries
A number of institutions and thinkers have produced helpful ‘prompt banks’ or ‘prompt libraries’ that offer example prompts for specific tasks. These can be a fantastic way to get started, and to start to learn what effective prompting can look like.
A couple of examples are as follows:
- The AI for Education Prompt Library offers a range of suggested prompts for applications by both educators and learners.
- The Wharton Generative AI Labs Prompt Library collects some very sophisticated prompts for tasks from creating simulations to meeting facilitation.
For more on possible applications of these tools, visit:
Applications for Student Learning
Assignments in the Context of AI
Labor-Saving for Educators