What "creating a GPT" actually means

Creating a GPT means building a customized version of ChatGPT that follows specific instructions, uses particular documents or data, and behaves in a consistent way for a defined purpose. You are not writing code or training a machine learning model from scratch. You are using OpenAI's GPT Builder tool to set up rules, upload files, and configure how an existing AI model responds to your prompts.

The result is a tool that remembers your preferences, stays focused on a narrow task, and can reference documents you provide without you having to paste them into every conversation. A lawyer might create a GPT that knows their firm's procedures and case templates. A teacher might create one that grades essays using a rubric they define. A small business owner might create one that answers customer questions using their FAQ document.

This is different from using ChatGPT directly. ChatGPT starts fresh each conversation with no memory of your preferences or your files. A custom GPT remembers both.

Key Takeaways

  • You need a ChatGPT Plus subscription (currently $20 per month) to access the GPT Builder tool; a free account cannot create custom GPTs.
  • The GPT Builder walks you through naming your GPT, writing instructions for how it should behave, and uploading files it should reference.
  • Custom GPTs run on the same underlying model as ChatGPT, so they have the same limitations and knowledge cutoff date as the standard version.
  • You can share your custom GPT with others via a link, or keep it private for your own use only.
  • Creating a custom GPT does not require coding knowledge, but the clearer your instructions, the more reliably it will behave the way you intend.

What you need before you start

You must have a ChatGPT Plus subscription. This costs $20 per month in the United States, though pricing varies by country. A free ChatGPT account cannot access the GPT Builder. If you have a free account, you can still use custom GPTs that other people have created and shared with you, but you cannot build your own.

You also need to know what you want your GPT to do. The clearer your purpose, the better the result. "A GPT that helps me write" is too broad. "A GPT that rewrites my blog posts in a conversational tone while keeping the same facts" is specific enough to work with. Spend a few minutes writing down what task you want this GPT to handle and what rules it should follow.

Finally, gather any documents, templates, or reference materials you want the GPT to use. These can be PDFs, text files, spreadsheets, or images. OpenAI's GPT Builder can handle most common file formats. You do not need to have these ready before you start — you can upload them during the setup process — but having them organized saves time.

How to access GPT Builder and start creating

Log into ChatGPT at chat.openai.com with your Plus account. On the left sidebar, you will see a menu. Look for the option labeled "Explore" or "My GPTs" (the exact label changes with updates). Click it, then select "Create a GPT" or the plus-sign button next to "My GPTs".

This opens the GPT Builder interface. You will see two panels: on the left is a conversation window where you can describe what you want, and on the right is a preview of your GPT as it takes shape. You do not have to use the conversation window — you can skip straight to the configuration form — but many people find it easier to describe their GPT in plain language and let the builder suggest settings.

Start by giving your GPT a name. This should be clear and specific: "Customer Support Bot" rather than "Helper". Then write a brief description of what it does. This description appears when you or others search for your GPT, so make it accurate.

Writing instructions that actually work

The instructions section is where you define how your GPT behaves. This is the most important part of the setup. Write clear, specific rules about what the GPT should do, what it should refuse to do, and how it should handle edge cases.

Good instructions are concrete. Instead of "Be helpful and friendly," write "When a user asks about shipping, tell them we ship within 3 business days to the continental US and 5 to 7 days to Alaska and Hawaii. If they ask about international shipping, say we do not currently ship outside the US." Instead of "Be professional," write "Use a formal tone. Do not use slang or emojis. Address the user as 'you' rather than by name unless they provide it."

You can also tell your GPT what it should not do. "Do not make up product prices. If you do not know the price, say so and direct the user to the pricing page." "Do not give medical information. If someone describes symptoms, recommend they see a doctor." These guardrails prevent your GPT from confidently stating things it should not.

Keep your instructions between 150 and 500 words. Longer instructions do not always work better — in fact, contradictory or overly detailed instructions can confuse the model. Test your GPT after writing instructions, and refine them based on how it actually behaves.

Uploading files and setting knowledge sources

In the GPT Builder, you will see a section for uploading files. This is where you add documents your GPT should reference. Click "Upload files" and select PDFs, text documents, spreadsheets, or images from your computer.

The GPT will read and remember these files for the duration of any conversation. If you upload a company handbook, the GPT can answer questions about your policies. If you upload a product catalog, it can describe your products. If you upload a list of FAQs, it can answer common questions without you having to type them in each time.

There is a file size limit — currently 20 MB per file and 10 files per GPT — so very large databases do not work. If you have a huge document, consider splitting it into sections or uploading only the most relevant parts.

You can also tell your GPT to search the web for current information. Toggle "Web browsing" on if you want your GPT to look up today's news, current prices, or recent events. Keep in mind that web browsing takes longer and uses more of your monthly message quota than a GPT that only uses uploaded files.

Testing and refining your GPT before sharing

Before you share your GPT with anyone, test it thoroughly. In the preview panel on the right, type sample questions and see how it responds. Ask it to do the exact task you designed it for. Ask it edge cases and tricky questions. Try to break it.

If the GPT gives wrong answers, contradicts itself, or ignores your instructions, go back and revise. You might need to rewrite your instructions to be more specific, upload additional files, or remove files that are confusing it. This is normal — even well-designed GPTs need tweaking.

Pay attention to whether the GPT is actually using your uploaded files or just relying on its general knowledge. Ask it a question that can only be answered by reading your documents. If it gives a generic answer instead, your files may not be clear enough, or your instructions may not be directing it to use them.

Sharing your GPT or keeping it private

Once you are satisfied with how your GPT works, you can choose who sees it. In the GPT Builder, look for sharing settings. Your options are usually:

  • Only me: The GPT is private. Only you can access it. This is useful for personal tools or internal business use.
  • Anyone with a link: The GPT is not listed publicly, but anyone you give the link to can use it. This works well for sharing with a team or a specific group.
  • Public (in the GPT Store): Your GPT appears in OpenAI's GPT Store, where any ChatGPT Plus user can find and use it. This is useful if you have built something genuinely useful and want a wider audience.

If you choose to make your GPT public, write a clear description and test it one more time. Public GPTs reflect on you or your organization, so accuracy and reliability matter.

What your custom GPT cannot do

A custom GPT is still built on the same underlying model as ChatGPT, so it has the same limitations. It cannot access real-time data unless you turn on web browsing. It cannot connect to your email, calendar, or other apps (though some integrations are being tested). It cannot perform actions in the real world — it can only give you information and suggestions.

It also has a knowledge cutoff date, meaning it does not know about events after a certain point in time. As of early 2024, ChatGPT's knowledge cuts off in April 2024, though this date shifts as OpenAI updates the model. If your GPT needs to know about recent events, web browsing is essential.

Finally, your GPT will occasionally make mistakes or "hallucinate" — confidently state things that are not true. This is a known limitation of large language models. You can reduce this risk by uploading accurate reference documents and writing instructions that tell your GPT to say "I do not know" rather than guess, but you cannot eliminate it entirely.

Frequently Asked Questions

Do I have to pay extra to create a custom GPT?

No. If you already have a ChatGPT Plus subscription ($20 per month), creating custom GPTs is included at no additional cost. You do not pay per GPT created or per conversation using a custom GPT.

Can I edit a custom GPT after I create it?

Yes. You can return to any GPT you created, open it in the GPT Builder, and change the name, description, instructions, or uploaded files. Changes take effect when ready for new conversations, though existing conversations may not reflect the updates.

What happens if I delete a custom GPT?

If you delete a custom GPT, it is gone permanently. If you shared it with others via a link, that link will stop working. If it was in the public GPT Store, it will be removed. Back up any important instructions or configurations before deleting.

Can I use a custom GPT to train it on my own data?

No. Uploading files to a custom GPT does not train or fine-tune the underlying model. The GPT reads your files during conversations but does not learn from them in a way that changes how it works for future conversations. If you need true fine-tuning, you would use OpenAI's API and fine-tuning service, which is a different process and requires coding knowledge.

What if my custom GPT gives wrong information?

Revise your instructions to be more specific, upload corrected reference documents, or remove files that contain errors. You can also add an instruction like "If you are not certain of the answer, say so and ask the user to verify." Test the changes before sharing the GPT with others.