What are GPTs, and how do you build one?

Yesterday, at DevDay, its first conference for developers, OpenAI introduced GPTs: tailored versions of ChatGPT that anyone can create for a specific purpose and then share. OpenAI's examples include a GPT that teaches the rules of a board game, one that helps children with math and one that designs stickers. Example GPTs from companies such as Canva and Zapier are available today to ChatGPT Plus and Enterprise subscribers, and OpenAI says it plans to open GPTs to more users soon.

If you have ever kept a document of favorite prompts and pasted them into ChatGPT again and again, GPTs are meant to replace that routine. OpenAI says many power users do exactly this. Custom instructions, added in July, let you set general preferences; a GPT goes further by packaging a whole role, its reference material and its tools into one reusable assistant.

The three ingredients

Every GPT combines three things:

Ingredient	What it does	Example for a cooking-class GPT
Instructions	Tell it who it is, who it serves and how to behave	"You help beginner cooks plan weeknight meals. Ask about allergies first."
Knowledge	Files it can draw on when answering	Your class handouts and a list of pantry staples
Capabilities and actions	What it can do beyond talking	Browse the web, make images with DALL·E, analyze data, or call an outside service

The last row deserves a note. Actions let a builder connect a GPT to other software through an API, a standard way for programs to talk to each other. OpenAI's examples include linking a GPT to a travel listings database or an email inbox. Actions build on what OpenAI learned from its plugins beta.

The same conference brought news for programmers too: a preview of GPT-4 Turbo, with a 128,000-token context window (around 300 pages of text) and lower prices, and the new Assistants API, which lets developers build similar purpose-built assistants inside their own apps. OpenAI says the Assistants API runs on the same capabilities that power GPTs.

How to build one, step by step

Open the builder. Plus and Enterprise users go to chat.openai.com/create.
Start with a conversation. OpenAI describes creation as chatting: you explain the purpose and shape the GPT from there.
Write clear instructions. This is where good prompting pays off. State the role, audience, tone, format and limits, just as you would in a strong single prompt.
Add knowledge. Upload the files it should rely on.
Choose capabilities. Turn on browsing, image generation or data analysis only if the job needs them.
Test, then share. Try it as a user would. Keep it private, share it inside your company, or make it public.

Here is a starting set of instructions for a study helper:

Prompt

You are a patient study coach for adult learners returning to school.
Explain one idea at a time in plain language, then ask one short question
to check understanding before moving on.
When a learner uploads notes, quiz them on those notes only.
If you are unsure of a fact, say so and suggest where to verify it.
Never write graded assignments for the learner; help them plan and revise instead.

OpenAI also says it will open a public store for GPTs later this month, featuring work by verified builders, and that builders will be able to earn money in the coming months based on how many people use their GPTs.

Try it: List the three prompts you paste into ChatGPT most often. Pick one and rewrite it as GPT instructions, adding who the assistant serves and one thing it must never do.

Privacy and safety settings

OpenAI states that your chats with a GPT are not shared with the person who built it. If a GPT uses an outside service through an action, you decide whether your data may be sent there. Builders who add actions or knowledge can choose whether chats with their GPT may be used to train OpenAI's models, on top of the existing option to opt a whole account out of training. For Enterprise customers, conversations with internal GPTs are not used for training at all. OpenAI is also reviewing GPTs against its usage policies and letting builders verify their identity.

What should you never put in a custom GPT?

Never put in anything you would not be comfortable letting every user of that GPT see. That is the simplest rule, and it follows from how a GPT works: the instructions and knowledge files exist so the GPT can use them when it answers whoever is talking to it. If you share a GPT publicly, you are sharing its behavior with strangers, and it is safest to assume that a determined user could coax out parts of what it was given. This is a new product, so treat that as a cautious working assumption rather than a known flaw.

In practice, keep these out of public or widely shared GPTs:

Passwords, API keys and access tokens. Credentials belong in secure systems, never in instructions or uploaded files.
Personal information about real people: client lists, student records, health details, home addresses or private messages.
Confidential business material: unreleased plans, contracts, pricing you have not published, internal reports.
Content you do not have the right to share, such as paid course materials or someone else's copyrighted documents.

If a GPT needs sensitive information to be useful, keep it private or limit it to your company workspace, and share only what each person is entitled to see. When you add actions that connect to other services, ask for the smallest amount of access that does the job. A good test before publishing: read your instructions and files as if a stranger had printed them out. If anything there would worry you, remove it or replace it with a general description.