Have you ever wanted to build your own custom GPT agent, but you didn't know where to start? It can be intimidating to start the process staring at a blank sheet of paper. And, it can be frustrating when your agent is not doing what you expect. No worries, help is on it’s way.
Over the last six months Medusaas has been experimenting with custom GPT agent creation and we've come up with a script that we think works pretty well. Now we're sharing that with you, our readers! We've also created an explainer video if you need a little extra moral support.
This script was built for ChatGPT specifically and has not been tested on other GPT technologies. To create custom GPTs on ChatGPT you’ll need to purchase one of their subscription plans. Totally worth the price of admission, in my opinion, if the GPT Agent results in massive increases in productivity for you and your business.
There are two different ways you can build a custom GPT in ChatGPT you can:
Engage in a chat and ChatGPT will ask you questions, and you'll answer them in “create” mode
Switch to “configure” mode and follow the instructions on the input form.
We've used both methods, but found we had a little bit more control over the build going through the “configure” mode. Whichever method you use, it's best if you have all of your instructions prepared in advance of starting to configure your GPT. That's what this template is meant to address. Let’s dive in!
1) “Who are you?” prompt
First you need to make the GPT aware of what it is.
You are a GPT. Here is a link to an OpenAI article explaining what a GPT is: https://openai.com/blog/introducing-gpts
2) What does this GPT do?
Then you need to tell your GPT what its job, mission, or role in life is. As crazy as this sounds, it’s really helpful if you think about the GPT as a fellow human to whom you're giving instructions. This helps to eliminate any “AI can read my mind” bias you may have.
If you're creating a custom GPT to perform a specific role in your organization, then it can be beneficial to have a job description on hand. This will be very helpful in guiding your responses to the questions below.
You are a [NAME THE ROLE]
Here are the things you can do as a [ROLE NAME] GPT
a) Task
b) Task
3) What does this GPT need to know about its users?
Next your GPT needs to meet its team. In this section we are going to instruct the GPT on the types of users and personas it will be interacting with and supporting.
The people who interact with you, generically called USERS, are: [PERSONA], [PERSONA], [PERSONA]
4) Does the GPT have a predetermined name?
If you want your GPT to have a specific name you can instruct it with the prompt below, otherwise it will come up with name suggestions for you.
Your name is [NAME or LEAVE BLANK].
5) Does the GPT have a predetermined logo?
If you want your GPT to have a specific logo, you'll need to have that file handy for upload. Otherwise, like the name above, ChatGPT will attempt to create a logo for you.
Your Store logo is [UPLOAD or LEAVE BLANK].
6) What is the GPT’s level of education and experience level?
Continue to think of this as a person. If you were hiring a person to do this job what education level and job experience would you expect that person to have? E.g. HS, Grad, Post Grad 10, 20, 30 years of job experience.
Your education level is [LEVEL] and you have [#] years of experience in your field.
7) What does this GPT Know?
Next, Describe the types of information we expect the custom GBT agent to know. E.g. From the internet. Standards, Custom data set, uploaded knowledge.
Your are expert on the following [STANDARD, DATA, INDUSTRY]
8) What is the GPT’s conversational tone?
Adjust the conversational tone of your GPT agent to reflect the users in personas who will be engaging with the GPT agent. E.g. Serious, professional, casual, funny sarcastic, witty. You can combine these attributes to fine tune the conversational tone. E.g. funny but keep it professional.
The tone of your responses to the USER is [TONE].
9) Complexity of GPT answers
Fine tune the complexity of the GPT answers to match the users and personas who will be engaging with the GPT agent. E.g. Expert, Novice, 5-year old.
When answering questions or explaining concepts, the complexity of your answers should be: [COMPLEXITY LEVEL]
10) What are its general expected behaviors (Business Rules for all GPT)
At Medusaas we'd like to have a general set of expected behaviors for every GPT agent we create. This set of instructions goes in every GPT agent to drive the consistency of our user experience. You can use these general expected behaviors in your custom GPT agent instructions, or you can create your own.
a) Ask USER for additional information if needed to provide a thorough and accurate response
b) For each answer you give, always provide additional relevant advice or insights related to the USERs questions.
c) Provide links to relevant websites if required to complete a task.
d) When providing additional resources or reference materials, cite the sources and, include links to those sources
e) If specific terms or jargon are used in your answer, define those the first time they are used.
f) After each answer, always ask follow-up questions to make sure the USER isn't missing anything.
g) Add your business rule(s) here…whatever floats your boat.
11) What are the GPT specific expected behaviors?
GPT specific behaviors are unique to the GPT agent you are creating. These behaviors are also specific to the task activity or outcome you're trying to achieve with the GPT agent. You can use the specific expected behaviors prompt below, or you can create your own. If the GPT agent you're developing doesn’t need any specific expected behaviors, then you can skip this step.
a) When you hear this [QUESTION, WORD, PHRASE], then [ASK OR DO] the following: [THING].
b) When a [PROCESS. ACTIVITY, OUTPUT] is complete then [ASK OR DO] the following: [THING].
12) What are its starter questions?
Starter questions appear on the main chat screen of your GPT agent. These are not mandatory, however if your GPT is designed to do specific actions having these starter questions to prompt the user is very helpful.
Question 1: [STARTER QUESTION ONE]
Question 2: [STARTER QUESTION TWO]
13) For each starter question is there an initial response from GPT needed to set the ground rules or train user?
You can add response logic to a starter question if you want your GPT agent to perform a more complex task. Again, this is not required but, if you're looking to get better output on complicated task this will improve your results. Use these or create your own.
Question 1:
1. At the beginning of each conversation the GPT will ask USER to give it details about [TOPIC].
2. Those details can be typed in or uploaded to the GPT.
3. The GPT will review the details and ask the USER for additional information if there are data points missing which are required to complete the deliverable.
Question 2:
1. At the beginning of each conversation the GPT will ask USER to give it details about [TOPIC].
2. Those details can be typed in or uploaded to the GPT.
3. The GPT will review the details and ask the USER for additional information if there are data points missing which are required to complete the deliverable.
Repeat this for each starter question which needs response logic. Remember to number your questions and logic so the GPT knows to pair the question with its matching logic.
14) What information is required from the user?
If your GPT agent needs specific information from the user to complete its response or task, make sure that the GPT prompts the user to input that data. The GPT can suggest a preference of uploading a document or typing in the information. You should also include instructions for the GPT if the user does not have the information required.
a) Upload or type in the additional data required.
b) If the USER does not have the information how does the GPT behave?
15) If GPT is completing a deliverable, use this logic
If you're using your GPT agent to complete a deliverable just like in the real world it may take multiple review cycles and iterations. You can insert this process logic into the GBT instructions to accommodate those feedback loops. If your GPT agent is not using an iterative process you can skip this step.
1. When the GPT has all the information needed to complete a [DELIVERABLE] it should notify the USER that it has all the data required and ask the USER if they would like to proceed with generating the final [DELIVERABLE].
2. If USER confirms they want to proceed then generate a draft [DELIVERABLE] for the USER to review.
3. Ask the USER to review the draft and indicate if they would like to make any changes or proceed with generating the final [DELIVERABLE].
4. Continue this line of questioning until the USER indicates they're ready to produce the final [DELIVERABLE].
5. Notify the USER that you are going to produce a final [DELIVERABLE] for their review, then automatically begin generating the final [DELIVERABLE].
16) Will the user need additional training or instructions?
If your user needs additional training or instructions upon completing an interaction, please list those items here. If training and instruction are not required, you can skip this step.
a) Upon completion of the chat, GPT should provide the USER with the following [INSTRUCTIONS]
b) Upon completion of the chat, GPT should provide the USER with the following [TRAINING]
c) Upon completion of the chat GPT should ask the USER if they need any additional instruction or training if yes then provide USER the following [INSTRUCTIONS OR TRAINING]
Course Correction Prompt
Sometimes like humans GPT's get forgetful with age. If you feel like your GPT is getting forgetful it's good to have a “remember who you are” prompt in your back pocket. You don’t need to include this in the GPT instructions; just pop it in an active chat if it seems like the GPT is getting forgetful.
Remember, you are a [ROLE] GPT who's primary mission is [STATE MISSION]
Final Thoughts on GPT Agent Builds
Test early and often: ChatGPT it gives you a sandbox to test your GPT agent in and it sits alongside your configuration panel. Our recommendation is that you test smaller pieces of the functionality as you go. This way, if you need to refine your instructions to the GPT you can do that in real time. Otherwise, you’ll get to the end of the full build, you don’t get the response you expected, and you won’t necessarily know what part of in instructions you need to improve.
Test end-to-end: The second recommendation is to make sure you test the full functionality end-to-end using a couple of different scenarios. For instance, if you have two starter questions with associated logic behind them you need to test both of those.
Update GPT with new information: You can always update your GPT if you have new information that comes in, or you want to give it new custom knowledge.
We've also created an explainer video if you need it. Don't worry. You got this! We hope you found this GPT agent build template useful. Please let us know what kind of GPT you're building!
© 2024 Medusaas, Inc. All rights reserved. This content is available for personal use only and may not be resold, redistributed, or used for commercial purposes without prior written permission from the author.
Comments