How to Create a Custom GPT Step by Step

To create a custom GPT, sign in to ChatGPT on a plan that includes custom GPTs (Go, Plus, Pro, Business or Enterprise), open the GPTs area from the sidebar, and start a new GPT. You then describe what it should do, write its instructions, upload knowledge files, choose which tools it can use, and publish it privately, by link or to your workspace. A simple, useful GPT takes about 20 to 30 minutes to build.

This step-by-step guide covers planning, writing instructions that actually hold, adding knowledge and actions, testing, sharing, and the mistakes that make most custom GPTs disappointing.

What is a custom GPT and when should you build one?

A custom GPT is a tailored assistant that lives inside ChatGPT. Instead of pasting the same long prompt every time, you save it once with supporting files, and anyone you share it with can use it with a click. OpenAI calls them GPTs, and they are separate from the underlying GPT-5.6 and GPT-6 models, which power them.

Build one when a task is repeatable, rule-based and needs the same context each time. Good examples:

• A brand voice writer that turns rough notes into on-brand LinkedIn posts
• A customer FAQ helper trained on your policies and product sheets
• A proposal drafter that follows your agency’s structure and pricing rules
• A study tutor that quizzes you on your own lecture notes
• An SEO brief generator that outputs headings, entities and questions in a fixed format

If you only need personal context for ongoing work, a project may be simpler. Our guide to using Claude Projects shows a similar workspace approach, and our ChatGPT vs Claude comparison explains which assistant suits which kind of work. New to ChatGPT entirely? Start with our complete ChatGPT beginner guide.

What you need before you start

Before you open the builder, prepare:

• A one-sentence purpose. “Turns meeting notes into client-ready follow-up emails in our house style.”
• Two or three example outputs you consider excellent.
• Reference files such as a style guide, FAQs, product sheet or process document, cleaned of anything confidential.
• Five test prompts, including a couple of awkward ones the GPT should refuse or handle carefully.

For plan details, our guide to ChatGPT pricing explains every tier, and our look at AI team plans helps if you are deciding whether Business seats are worth it for sharing GPTs internally.

How to create a custom GPT step by step

OpenAI updates the builder often, so the steps below describe where things usually are rather than exact button labels.

1. Open the GPT builder. In ChatGPT on the web, open the sidebar and go to the GPTs area. Choose the option to create a new GPT. The builder usually offers two ways to work: a chat-style assistant that asks you questions, and a manual configuration view with fields you fill in yourself.
2. Sketch it with the chat builder (optional). Describe your GPT in plain language, such as “A GPT that writes Instagram captions for a Jaipur jewelry brand in a warm, premium voice.” The builder drafts a name, description and instructions. This is a fast starting point, but always review what it writes.
3. Switch to manual configuration. Open the configuration view and edit each field directly. This is where you get precise control. Add a clear name, a one-line description that tells users what the GPT does, and a profile image if you like.
4. Write the instructions. Paste in a structured brief (template in the next section). This is the most important step. Be explicit about the role, steps, output format and what the GPT must not do.
5. Add conversation starters. Write three or four example prompts that appear as buttons for users, like “Turn these notes into a follow-up email.” They teach people how to use the GPT correctly.
6. Upload knowledge files. Add your reference documents in the knowledge section. The GPT can search these when answering. Use clean, well-structured files with clear headings.
7. Choose capabilities. Turn on only the tools the GPT needs, such as web search, image creation, or code and data analysis for spreadsheets. Fewer tools mean fewer surprises.
8. Add actions if needed. Actions let a GPT call an external API, such as your CRM or a weather service. You provide an API schema and authentication details. Skip this for your first GPT unless live data is essential.
9. Test in the preview panel. Run your five test prompts in the preview side of the builder. Adjust instructions after each failure and test again.
10. Publish and share. Save the GPT and choose who can access it: only you, anyone with the link, or (on Business and Enterprise) people in your workspace. Public listing in OpenAI’s GPT directory may also be available depending on your account.

How to write custom GPT instructions that actually work

Most weak GPTs fail because the instructions are either too vague (“be helpful and creative”) or a wall of text with conflicting rules. Use a short, structured template:

ROLE: You are [role] for [business or audience].
GOAL: Help users [specific outcome].
PROCESS:
1. If key details are missing, ask up to 3 short questions.
2. Check the knowledge files before answering.
3. Produce the output in the format below.
FORMAT: [length, headings, bullets, tone, spelling].
RULES:
- Use only facts from the knowledge files for prices and policies.
- If the answer is not in the files, say so. Do not guess.
- Never [off-limits topics or actions].
EXAMPLES: [one short good example].

Three principles make instructions stick:

1. Put the most important rules first. Instructions near the top tend to be followed more reliably.
2. Use numbered steps for process. Models follow ordered steps better than flowing paragraphs.
3. Define the failure behavior. Telling the GPT what to do when information is missing prevents invented answers.

Our prompt engineering guide explains these techniques in depth, and our list of ChatGPT prompts for marketing is a good source of task ideas to turn into GPTs.

Knowledge files, capabilities and actions explained

Knowledge files

Knowledge files are documents your GPT can search when it answers. They are ideal for FAQs, policies, product catalogs, style guides and course notes. Tips for better retrieval:

• Use clear headings and short sections so relevant passages are easy to find.
• Remove duplicates and outdated versions.
• Prefer text-based formats over scanned images.
• Tell the GPT in its instructions which file to use for what, for example “Use pricing.pdf for all prices.”

Capabilities

Capabilities are built-in ChatGPT tools you switch on or off per GPT. Web search helps with current information. Image creation suits design or social GPTs. Code and data analysis lets the GPT work with spreadsheets and CSV files, which is useful for reporting GPTs; see our guide to ChatGPT for Excel and data analysis for ideas.

Actions

Actions connect a GPT to external services through an API. For example, a sales GPT could look up a lead in your CRM, or a support GPT could check order status. You describe the API with a schema and set up authentication. Actions are powerful but add security and maintenance work. For no-code alternatives that connect apps together, our guide to automating tasks with Zapier and AI may be simpler.

Custom GPT ideas by user type

If you want a public-facing chatbot on your own site rather than inside ChatGPT, a GPT is not the right tool. See our guide on adding an AI chatbot to your website instead. Small business owners can find more use cases in our AI prompts for small business owners.

Common mistakes when building custom GPTs

One more mistake: expecting a GPT to be perfectly consistent. It still runs on a language model, so outputs vary and can contain errors. Keep a human review step for anything client-facing, and write a simple team policy with our AI usage policy template.

Frequently asked questions

Next steps: build one small GPT this week

Pick the single task you repeat most, write a one-sentence purpose, collect two good examples and one clean reference file, and build a focused GPT in the manual configuration view using the template above. Test it with five real prompts, fix the instructions after each failure, and share it only when it passes. Once one GPT saves you real time, build the next.

For official details, see OpenAI’s ChatGPT pricing page and the ChatGPT Business overview. Plan features and prices change often, so confirm current details on OpenAI’s site before you upgrade.