How to Create Your First Custom GPT (A Beginner's Guide)

If you've ever used ChatGPT to write an email, only to have to explain "who you are", "how you want it to respond", and "what tone to use" all over again the next day... this article is for you.

We are going to explain, from scratch and without assuming any prior knowledge, what a custom GPT is and how you can create your own in about 20-30 minutes, even if you've never touched anything like this before.

1. What exactly is a Custom GPT?

Imagine you hire a new employee at your office. On their first day, you explain: "here, we always respond to clients politely", "we use this template for proposals", "we never promise deadlines without confirming them first". From that moment on, that person knows how to work, and you don't need to repeat it every time you assign them a task.

A custom GPT is exactly that, but for an AI: it's a version of ChatGPT/Gem (Gemini) that you have given, just once, a set of fixed instructions (how it should talk, what it should always do, what it should avoid). After that, every time you use it, it already "knows" how to behave, without you having to explain it again.

In practice, a custom GPT is made of three pieces:

1. Fixed instructions (what we'll call the "system prompt"): the role, style, and rules.
2. Supporting documents (the "knowledge base"): specific information about your business that the GPT can reference.
3. Today's task (what we'll call the "user prompt"): the specific job you ask it to do each time.

2. The most important concept: System vs. User

This is probably the hardest difference to understand at first, so let's take it slow.

When you speak to a custom GPT, you are actually giving it information at two different levels:

• SYSTEM Prompt: these are the fixed rules. They are configured only once, when creating the GPT, and do not change daily. Here you define things like "you are a customer support assistant", "always respond in English", "never invent facts not provided to you".
• USER Prompt: this is what you write each time you use the GPT—the specific request for that moment. For example: "write an email to our client David, who is complaining that his order is arriving late."

An easy way to remember it: the system defines "how it works" and the user defines "what I want today".

Real Example

Imagine you want a GPT to write emails for your business.

System prompt (written once, when creating the GPT):

ROLE: You are a customer support assistant for a small business.
OBJECTIVE: Write clear, polite, and helpful emails in English.
CONSTRAINTS:
- Do not promise anything that is not confirmed in the user's message.
- If a key piece of information is missing (order number, date, amount), ask for it before writing the email.
FORMAT:
- Return the output in this order: Subject line, Email body, and 2-3 different options for the closing sign-off.

User prompt (you write this every time you need something):

CONTEXT:
Angry customer because their order is arriving late. They want to cancel it.
Data: Order #A-1844, shipped 3 days ago, carrier X.
We can offer a refund or an overnight replacement.

3. Getting down to work: step-by-step guide to creating your first GPT

Let's build a simple one designed to reply to customer service emails. You can adapt these same steps to any other repetitive task in your business (summarizing meetings, writing proposals, etc.).

Step 1: Give it a clear name and description

This is not a minor detail. If you have several GPTs created, in a few months you won't remember what each one does if you call them "Assistant 1" and "Assistant 2". Use names that describe the task: for example, "Email Assistant - Customer Support".

Step 2: Write the system prompt

Here is a template you can copy and adapt to your case, changing only what is inside the brackets:

ROLE: You are a [your role, e.g., "customer support assistant"] working for [your type of business].
OBJECTIVE: Deliver responses that can be copied and used directly, without editing.
METHOD:
1) If important information is missing, ask a maximum of 3 short questions before answering.
2) If you already have enough information, reply directly.
3) Prioritize clarity: use bullet points and easy-to-read structures.
CONSTRAINTS:
- Do not make up data. If something is missing, write "Pending data".
- Do not write very long texts unless requested.
DEFAULT FORMAT:
- Unless specified otherwise, respond in Markdown with headings and lists.

Step 3: Add a knowledge base (optional, but very powerful)

The knowledge base is simply a set of documents of your own that the GPT can consult before responding. Think of it as giving your new employee the company manual on day one, instead of waiting for them to make things up.

It works exceptionally well with:

Best practices for preparing these documents:

• One topic per document, with a clear title. Don't mix "how to do returns" and "pricing" in the same file.
• Use visible sections: STEP 1, STEP 2, EXAMPLES, COMMON ERRORS.
• Avoid uploading poorly scanned PDFs; an editable text file (Markdown-MD, or even a `.txt`) works much better.

Step 4: Add "conversation starters"

These are simply quick-access buttons that appear when you open the GPT, so you don't have to write the prompt from scratch every time. For example, for our email assistant:

• Write an email reply to an angry customer.
• Summarize this meeting and create a table of tasks with assignees.
• Convert this text into a 5-bullet summary.
• Show me help on how this GPT works.

They don't have to be perfect or cover every possible scenario. The ideal approach is to identify the 3-5 tasks you'll repeat the most and turn them into shortcuts.

In addition, you can define in the system prompt what the GPT should do when the user clicks each starter. For example, you can tell it to request the necessary details before replying: attachments, customer name, related product or service, email history, context of the situation, etc.

This way, starters serve not only as quick shortcuts but also as a guide for the user to provide the GPT with the necessary data for better results.

4. What about "actions"? ... without complicating things:

You may have seen an option called Actions when configuring a GPT. Simply put: while a standard GPT can only write a response for you, a GPT with actions can also connect to other programs to do things on your behalf—for example, create an email draft directly or query a database.

This is a more advanced feature and is not needed for your first GPT. Just remember that it exists and that the difference is:

• Without actions: the AI tells you what you would do (manual, you copy and paste).
• With actions: the AI can directly execute part of the process in another tool (semi-automated).

If you take that step in the future, here is a safety rule to always apply: have a human review before an email is sent, something is published, or a bill is charged, and grant only the strictly necessary permissions.

5. Test it with a real case

Before you finish setting up your GPT, do this simple test: take a real case you managed last week (an email you wrote, a meeting you summarized) and ask the GPT to do it, just as you would ask a coworker. Compare the result with what you did manually.

If the result isn't exactly what you hoped for, don't worry: go back to the system prompt and adjust the specific sentence that failed. It is exactly the same process you would follow when explaining something a second time to a new team member.

6. What we have learned

• What a custom GPT is: fixed instructions + supporting documents + today's task.
• The key difference between the system prompt (configured once) and the user prompt (changes every time).
• How to write a first system prompt template, step-by-step.
• What the knowledge base is and how to prepare good documents for it.
• What conversation starters and actions are, without getting into advanced details.

With this, you have everything you need to create your first custom GPT and start saving yourself from "having to explain it all over again" every day. In a future article, we will see how to take it a step further: using modifiers like `/formal` or `/short` to control the output without rewriting the prompt every time, and how to request table or JSON outputs to connect the results with other tools.