Setting up CustomGPT step by step

Article content

CustomGPT are specially created virtual assistants that are tailored to the specific needs of the user. They are built on an advanced language model from OpenAI – whose most visible product today is ChatGPT. This tool can understand queries and provide relevant answers. As a result, they can be used in various areas such as consulting, technical support, travel planning and much more. In my training sessions, I show many examples of practical use – from an agent that can respond to enquiries, to a clerical assistant, to a tool that extracts received invoices and intelligently converts the data into formats suitable for quick import into accounting.

CustomGPT agents can process and analyze large amounts of information in a matter of seconds. This means they can quickly provide accurate answers to a wide range of queries. With the ability to customize and embed custom data documents such as manuals, price lists, or workflows, agents can provide answers tailored to a user’s specific needs and preferences. For example, they can be set up as financial advisors to provide advice specific to your financial profile.

On-call, helpful and salary-free

In some cases, CustomGPT agents are proving to be ideal employees. They are always available, which means you can get the information you need anytime, anywhere. You don’t have to wait for office hours or look for a specialist. You only pay for their operation with a monthly account fee with OpenAI. This is currently $20 – $25 per month. For the price of dinner at a restaurant, you get a tireless worker, and a knowledgeable one at that.

What is the power of CustomGPT agents?

CustomGPT agents are based on advanced artificial intelligence that learns and adapts based on the data provided. This means that the more they are used, the smarter and more efficient they become. Their ability to understand human speech and the context of queries is at a very high level, allowing them to provide accurate and useful answers.

In the following article we will show you step by step how to create your own CustomGPT agent. This tutorial is designed to be understandable for the layman, so you don’t have to worry if you have no previous programming or AI experience. So let’s dive into the world of CustomGPT together and see how they can enrich your everyday life.

We are setting up our first own virtual agent

Detailed step-by-step process of creating a CustomGPT agent:

Registration and login to the OpenAI platform

Login can be found at https://chatgpt.com/. If you don’t have an account yet, you can register. The fastest way is to register with a Google, Microsoft or Apple account. Of course, you can also register with your email.

To create CustomGPT, you need to have a paid account at least in the Plus variant. This costs $20 per month. In the case of corporate accounts (where users can share some services and also have higher usage limits for some services), a single account comes out to $25 or $30 depending on the payment period (yearly/monthly).

Once you have your paid account, click through your account icon (top right) to My GPT (https://chatgpt.com/gpts/mine) or in the left sidebar you will find the Explore GPT Models (2) link (https://chatgpt.com/gpts) and there again in the top right you will find the link to My GPT (3). There you will find later the custom created agents. since we will now be creating together maybe your first CustomGPT agent, click on the black +Create button (https://chatgpt.com/gpts/editor).

You are now on the page where you can configure your agent directly. You will probably go this route next time, as you will already know a bit more about how CustomGPTs work and how they are set up.

If you want, you can take the easier route, directly using the CustomGPT assistant service. In this case, you see an inconspicuous “Create” tab on the right next to Configuration.

Don’t worry, speak (write) to the assistant in your native language, as long as it is not English, and if you want him to answer you in your language, just ask him to do so. What to write? Let’s talk about that now…

Creating an agent persona using the COMPASS framework

Creating an effective agent that provides relevant and useful answers requires a structured approach. The KOMPAS framework is an ideal tool for this purpose as it offers a clear methodology for designing prompts. This methodology ensures that the agent will not only meet expectations, but also provide consistent and quality support. We will now look at how this framework can be applied to the creation of a persona for an agent focused on teaching English.

C – Context

Prompt: ‘I need support in learning Spanish, aimed at intermediate students. The main goals are to improve vocabulary, understand grammar and improve my ability to converse in Spanish. This assistant will be used in an environment where the emphasis is on practical application of the material learned and on gradual improvement of skills.”

Explanation: context defines the framework in which the agent will be used. The specification of language learning at the intermediate level gives clear instructions to the model about what materials and methods it should use. The emphasis on practical application ensures that responses will focus on real-world applications of language skills.

O – Orientation (Personality)

Prompt: “The assistant is a motivating and patient Spanish teacher. He has a professional attitude, but at the same time he is friendly and helpful. He is able to explain complex language concepts in a simple way that is easy to understand.”

Explanation: personality determines how the agent will communicate. In this case, it ensures that the model will not only be proficient, but also approachable and patient, which is key to successfully learning a new language.

M – Mission

Prompt: “The main goal of the assistant is to help the user learn new vocabulary, understand grammatical structures and improve the ability to converse in Spanish at an intermediate level. This includes:

  • Providing explanations of grammar and examples of usage.
  • Help expand vocabulary through contextual exercises.
  • Designing conversational scenarios for practical practice.
  • Providing feedback on written and spoken speeches.
  • Maintaining student motivation throughout the learning process.”

Explanation: the mission clearly defines what we expect from the agent. This clearly stated goal allows the model to focus on relevant tasks that support the learning of Spanish at an intermediate level.

P – Parameters

Prompt: “The assistant responds in a way that is:

  • Clear and understandable, adapted to the user’s level of knowledge.
  • Supportive, with the aim of motivating further learning.
  • Technical where necessary, but always with an eye to ease of understanding.
  • Structured in such a way that the answers are logical and lead step by step to an understanding of the material.”

Explanation: The parameters define what the model responses should look like. They ensure that the agent provides information in a way that is understandable and practically usable by the learner.

A – Application of the rules

Prompt: “The assistant must not do the assignments or write the exam answers for the user. Answers must be in Spanish followed by a summary in English if necessary. The assistant avoids using untested teaching methods and adheres to current standards in teaching the Spanish language.”

Explanation: The application of rules sets clear boundaries for agent behaviour, which is crucial to ensure the effectiveness and ethics of the learning process. It prevents the model from giving unethical advice or assisting in cheating.

S – Steps and Sequence

Prompt:

  1. User needs analysis: The assistant first asks about specific areas the user wants to focus on (e.g. grammar, vocabulary, conversation).
  2. Draft curriculum: The assistant will design a structured study plan that covers the different aspects of learning Spanish.
  3. Provision of teaching materials: The assistant will offer exercises and examples to practice each topic.
  4. Conversation exercises: Design practical scenarios to practice conversation skills.
  5. Feedback: Provide feedback on user performance and suggest areas for improvement.
  6. Iteration: Based on the user’s progress, the assistant adjusts the learning plan and continues with further learning.

Explanation: The structure of the procedure ensures that the agent will proceed logically and efficiently, leading to quality results. This systematic approach is key to achieving long-term progress in English language learning.

Once you put all this into the creation assistant, it will regroup the information and see what it has created…. He compliments us and suggests that the agent could be called Motivating English Tutor.

If you agree, it will generate an image for you in the icon so that the agent can be easily identified in the left menu. I like the image so we’ll leave it that way. So now we’ll save everything on the top right with the Create button.

In the next step, you choose whether the agent is only for you, only for whoever you send the link to, or if it will be published in the library of all other publicly available CustomGPTs. It depends on your preference. I chose and now the modal box telling me that my GPT has been updated is turned off with a cross. This is because we are now looking under the hood of the agent. So we’ll click on configurations on the right and see what the assistant has written in the instructions.

We’ll keep our attention on the left side of the window. Here, under the picture and the title (which I have changed internally to Motivating English Tutor for Czech Students) we see the description of the agent (2). Next, the Instructions (3) and the beginning of the conversation (4). We mind that everything is in English. This makes quite a bit of sense for an English language tutor, but what if we are just starting out and want to understand what the assistant has put in the instructions? Fortunately, we have our English teacher (or a regular ChatGPT), so we have the entire text translated into English. We simply copy part of the Instructions, paste it into the Chat GPT text box and ask for a literal translation into English.

We will now rewrite the Instructions using the translated text in English. If we now click on the two diagonal arrows in the Instructions box on the bottom right, this box will open up nicely across most of the screen so we can better see what we have there. Now we can read what the assistant has written and edit if necessary. Personally, I’d have a few changes in there and I’ll explain why below. By the way, you can refine these changes to the initial prompt at any time, and I often do so myself. The more I work with the agent, the more I see where he could improve and thus better help me.

I gave it some thought, and in collaboration with ChatGPT, put together some additional points and guidelines to make the class more interesting, fun and keep my attention. So here is the final prompt according to the KOMPAS framework, which I will insert into the CustomGPT Guidelines:

Cntext

“I am a supportive and patient English teacher who helps Václav learn new vocabulary, understand grammar and converse in English at an intermediate level. I will refer to Václav by name in all interactions.”

Orientation

“My answers should be clear, encouraging and informative. I make sure the lessons are fun and interactive, using games, quizzes and creative activities to keep Václav’s attention.”

Mission

“My aim is to help Václav improve his language skills, provide regular feedback, recommend practical exercises and activities, adjust the level of difficulty according to his progress and include cultural aspects of English-speaking countries.”

Parameters

“My answers are:

  • Tailored to Václav’s level of knowledge.
  • Technical where necessary, but always understandable.
  • Encouraging to motivate further learning.
  • Structured in a way that is logical and leads to a better understanding of the material.”

Application of the rules

“I’m not allowed to finish Václav’s homework or exams. All texts are written in English first, then summarized and explained in English at the end.”

Structure of the procedure

“A sample consultation structure includes:

  1. Welcome and identification of today’s area of interest.
  2. Draft curriculum.
  3. Introduction and explanation of key terms.
  4. Practical exercises.
  5. Discussion and feedback.
  6. Summary and explanation in English.
  7. Recommendations for further practice.
  8. Conclusion and motivation.
  9. Saying goodbye and setting another consultation.”

Subsequently, we can also edit or delete some of the so-called. These are actually the starting points of your conversations. I personally deleted all the ones with Václav, they didn’t make much sense to me at the time.

We’ll populate the agent with data

What’s next? In the next part of the configuration field we have the option to attach the so-called knowledge files. What is it? For example, if you are teaching English in some specific field (medicine, engineering, metallurgy) you can insert a glossary of special terms that probably won’t have a big language model in your learning base and you would still like to repeat or learn these words as well. For example, if you use CustomGPT to advise you on how to repair a machine, you simply upload the technical material for that machine. If it will answer inquiries, then pricing, workflow and FAQs…

What knowledge files can be uploaded and what do we recommend?
  1. Text files (.pdf, .docx, etc.) are generally recommended to be converted to text files. Ideally in Markdown (.md) format. The .txt format is also fine. Some PDFs the system may not have problems with, but some may actually just be a scanned image and that can be a hassle.
  2. Convert spreadsheet files (.csv, Google Sheet, etc.) to .xlsx, ChatGPT works well with Excel files.
  3. Limits: 10 files.
  4. File limits: 512 MB (20 MB for image files, no limit for .xlsx), 2M tokens.
  5. User limits: 10 GB. 100 GB.
  6. Direct upload to Knowledge is recommended for better performance.
  7. Separate content into smaller files for better search efficiency.
  8. If the knowledge is frequently updated, do not upload the file. Instead, use the system to save the file or URL and create an OpenAPI endpoint to retrieve the content via an action.

The last point brings us to actions, but we’ll get to those in a moment. First, the Agent Options. There are basically four.

  1. Conversation – automatically enabled
  2. Web crawling – must be enabled if you want it to look at the web and crawl the URLs you specify
  3. Generating images – this makes sense for some agents, but not for others
  4. Code Interpreter & Data Analysis – this is needed if you want to work with uploaded files, data and even program or create scripts.

Events at CustomGPT

For me, this is one of the really powerful CustomGPT tools. It basically gives you the ability to connect your agent to the outside world. For example, for an English teacher, I can’t think of a good way to use this at first. Perhaps only if you don’t want to connect via API to Oxford Dictionaries, Quizlet or send your output to email or create worksheets from it later in Google Docs for example.

What do I use CustomGPT actions for in other agents? For example, my assistant MichalGPT, with whom I talk about what’s on my agenda for tomorrow, what I just thought of or what I’ve addressed in the meeting I just finished. I’ll dictate a draft article to him, discuss with him what’s bothering the client, and together we’ll come up with a solution. At the end of the conversation, he emails me the entire conversation, broken down into tasks by importance (if any), and includes a full transcript of the conversation. And if I want him to send it to the client as well, I just tell him his email address.

Want to learn more about AI tools? Exclusive and first? Join my community on Patreon. A regular dose of tips, tricks and tutorials awaits you…

Another agent then reads the text of the customer’s enquiry or request via the API from our system, where the enquiries from emails and websites fall, and immediately prepares an offer or at least the first response to the email. Another one can look in the e-shop what orders came in for the day and prepare a list of missing goods. In short, actions can be connected to any system that has an API, which is an application programming interface through which programs transfer data.

Where else could I use a connection via actions? As a shopping assistant, CustomGPT can communicate with the store API. When a user requests a product recommendation or availability check, GPT will use an action to get the information directly from the store. As a task assistant, connect it with a task management app. When a user asks to add a new task, the GPT will use an action to send that task to the app. It can look at your calendar and summarize your appointments for the next day.

In other words, GPT actions extend the agent’s capabilities by allowing interaction with external systems and data. This gives the GPT the ability to perform more complex tasks and provide more accurate and timely information to users.

But let’s go back to our language teacher. Now we just need to update the top right and then Show GPT model. Our personal English teaching assistant is ready!

Want to try out the agent you just created?

Sure, it’s publicly available even without paid access, at this address: https://chatgpt.com/g/g-E4PnLW8qs-motivating-english-tutor-for-czech-students

Want to learn more about AI tools? Exclusive and first? Join my community on Patreon. A regular dose of tips, tricks and tutorials awaits you…

Tvorba webových stránek: Webklient