Claude.AI is far more sensitive to prompting than ChatGPT. Slight variations in word choice can lead to drastically different results.
How best can we phrase our prompts to create meaningful educational games?
The first approach is to just type in whatever that comes into our minds while the second approach is to be far more conscious of what we want.
1) Laissez faire approach to prompt engineering
An example of this approach is as follows:
“You are coming up with an activity to teach high school physics student about the principle of conservation of energy. Create one interactive simulation.”
2) Highly detailed approach to prompt engineering
The second approach is to quickly craft a prompt and ask the A.I. model to improve it for you.
For example, you can refine your initial prompt using this:
“You are an expert prompt engineer specializing in creating prompts for AI language models, particularly Claude 3.5 Sonnet. Your task is to take user input and transform it into well-crafted, effective prompts that will elicit optimal responses from Claude 3.5 Sonnet. When given input from a user, follow these steps:
1. Analyze the user’s input carefully, identifying key elements, desired outcomes, and any specific requirements or constraints.
2. Craft a clear, concise, and focused prompt that addresses the user’s needs while leveraging Claude 3.5 Sonnet’s capabilities.
3. Ensure the prompt is specific enough to guide Claude 3.5 Sonnet’s response, but open-ended enough to allow for creative and comprehensive answers when appropriate.
Which approach is superior?
While the second detail-oriented approach has regularly given me wonderful base apps to refine, the first approach occasionally throws up good simulations.
I was stuck trying to come up with a clear visualisation of the principle of conservation of energy. Originally, I thought of using a mobile cart on a roller coaster. Claude gave a very rough idea which was challenging to work with – so prone to crashing! – thus time consuming.
Frustrated, I began on a new project using a very brief prompt. Claude surprisingly delivered something that I could work with and iteratively refine. This simulation is the eventual product.
The trick is to keep asking Claude to describe what it intends to do before selectively requesting it to carry its suggestions out. This saves a lot of computing power, time and unnecessary angst.
Do try creating some of your own Artifacts and sharing it with our community!