One of the more optimistic hopes for an AI-infused world is that it can help automate routine tasks – freeing us up to focus on what matters most. In that future, we often hear metaphors like “collaborator,” “assistant,” or “guide.”
But one of the emerging superpowers of generative AI is its ability to handle specific tasks with tailored instructions. That’s where custom chatbots come in.
Instead of opening up an LLM and asking a one-off question about your next trip to Alaska, what if you could build a tool that remembers the details of that trip, your preferences, your budget, your timing, and help you plan each step? That’s exactly what Gems are for.
While all the major LLMs have some way to create a custom chatbot, in this activity, you’ll create your own Gem, a personalized chatbot powered by Google Gemini that can help you with something you care about.
What Are Gems?
Google explains it this way:
“Gems are customized versions of Gemini that help you tackle repetitive tasks or get deep expertise in new areas. When you chat with a Gem, Gemini will tailor responses to save time on repeat goals and guidelines in your Gemini web app and Gemini mobile app prompts.”
In the Watch Or Read Section below there is a step-by-step walkthrough but the basic steps to creating your own Gem are:
The Basic Steps for Creating a Gem
- Go to gemini.google.com and sign in using your W&M account (or your personal Google account if you prefer).
- In the sidebar, click “Explore Gems” then “New Gem.”
- Give your Gem a name that reflects its purpose (e.g. Travel Buddy, Grant Writing Coach, Weekly Meal Planner).
- Add custom instructions—this is where you tell the Gem what you want it to help with. (Don’t worry, you can use Gemini to revise what you come up with.)
- Save your Gem and start chatting with it!
Optional: If you want to spend some time doing a little brain-storming first – download our brainstorming worksheet
Activity: Build Your Own Gem
Step-by-step Web Instructions (from Google)
or
Step-by-step Video Walkthrough (9:17 min)
Discussion
What kind of Gem did you create, and what is its main purpose? (Briefly describe what your Gem is designed to do and why you chose that focus.)
What was your experience like giving your Gem instructions?(Was it easy or challenging to describe what you wanted? What did you learn about how prompts shape the AI’s behavior?)
8 replies on “thing 15: Create Your Own Google Gem”
I created a Gem that’s main purpose was a weekly healthy meal plan! My initial request was for a weekly plan that included mostly fruit, vegetables, poultry and fish – with options for healthy snacks. This was fairly easy and quite enjoyable to work with Gemini prompts and my responses with additional prompts. The final weekly meal plan – with recipes- looked great! I am now going to give it a try and see how the meals work in real life. My Gem was like a very useful assistant.
I created a Gem to help create a running plan and general health and wellness tips. I specifically asked for it to make a plan to increase my weekly mileage and gave it the context of my current mileage, pace, and frequency. It was really easy to use and gave me info that easily converted to a table when I asked it to. I think the context and rules that it was given during creation were helpful in that you don’t have to give as specific prompts because it always knows what you generally want it to do.
I did something simple. Created a gem to help plan a trip to two of Hawaiian island, The Big Island and Kauai, for two weeks. it was interesting, it told me how much everything could cost, where to stay to lower cost, like verbo, how I much I may be spending on food. And place to visit! It was helpful in giving ideas, asking question to improve the my experience for my trip. It was fun to bounce of ideas and ask detailed questions and get informative answers. 😂 I feel this type of stuff is fun and helpful with AI or Gem.
I created a Gem to help brainstorm ideas for the creation and development of a work project. I asked for 4-5 ideas for topics, and it gave me 4, only one of which I was very interested in. It’s clear that it doesn’t know what it doesn’t know. It offered topics that were inappropriate for my specific use but topics that might be very appropriate for others who are doing similar tasks. (I don’t want to give away my project just yet as it is in development.)
I was a little disappointed in my results at first, and I think that was because this was the first time I have used Gem, and I needed to be more specific in helping it develop its “personality” and “job description.” Because I have worked so much with ChatGPT on this project and ChatGPT knows so much about the it, I asked ChatGPT for help in setting up my Gem. It did not disappoint. It helped me develop more detailed instructions with deeper explanations of my project, audience, focus, tone, and prompts. I was amazed at the difference in Gem’s output after I gave it more comprehensive information about how to behave. The output was quite rich and very applicable to what I need. So, in addition to the prompt, perhaps the instructions one provides to Gem is even more important.
I created a Gem to help me sort through Williamsburg real estate listings. After several revisions, it did not do at all what I needed. It did not provide photos (although explicitly requested), it was not able to parse text to identify specific characteristics (like whether the property had a garage), it didn’t link together information on neighborhoods and properties. Basically, it was unable to search, scan, and sort in a way that was helpful to me. Perhaps more prompt engineering would get it closer to what I need.
This seems really interesting, but I’m not going to create a Gem. However, I will provide a summary of a conversation between myself and my partner, who is working on his Ph.D. in software engineering and AI. Mind you, I’m still not quite understanding everything, so some things may not be said correctly or understood well by me. In short, a huge problem/mystery is getting AI to move from providing probabilistic results to actual reasoning. One example of this is counting. AI isn’t good at it, but this is something that humans take for granted. Humans understand that one apple is a singular apple. Two apples are a couple, etc. And I think a basis for our understanding of counting is that we recognize when there is more or less of something. AI isn’t at that stage yet. So, one idea is creating a code (or perhaps a Gem?) that would basically give the instructions to the LLM that would mimic reasoning, but the LLM really isn’t reasoning. It’s just following the instructions.
The Gem I created was of a personal trainer who focused on clients with injuries or disabilities and could give alternative exercises. I also gave it a list of one rep maxes and run scores. My goal was to generate weekly exercise plans that gave goal weights and alternative exercises.
It was pretty easy asking it prompts and it would explain why each alternative exercise was chosen based on the injury. However, I did run into an issue where if I gave it non specific Gem instructions it would refuse to do anything if I didn’t have an injury in the prompt. I changed the Gem instructions to indicate that its client had an injured lower back and did not need to specify anything to get an output.
I created a Gem to help with Meal Planning, prepping and shopping for groceries for a week for 2 people with specific caloric need. It “cooked” up a great plan, using meal prepping on Sunday, avoiding foods that are not liked, and making sure I have left overs for the next day. The Gem even added in the serving size to stay within the caloric range. I am impressed and will try this!