How to create a custom GPT (step by step)

By Ankur Shrestha, founder of Twinsona – Updated July 2026

Creating a Custom GPT means walking through six fields inside ChatGPT: open the GPT builder on a paid plan, name it, write its instructions, attach your knowledge files, switch on the abilities it needs, then test it in the preview and save it as private, link-only, or public in the GPT Store. No code, and most people finish in under an hour. This guide runs each step in order, flags the limits that bite once real people start using it, and shows where an owned twin is the better tool.

The short version: Six steps build a Custom GPT inside ChatGPT: open the builder, name it, write instructions, attach knowledge files, enable abilities, then test and save. The build is quick and suits narrow, repeatable tasks. What it won't do is prove where an answer came from, live anywhere but inside ChatGPT, or let you price access on your own terms. When the thing wears your name in public, an owned AI twin is the better shape: it replies only from your own material, footnotes each claim, and leaves the audience and the earnings with you.

What you need before you start

You need a paid ChatGPT plan, because the GPT builder is not on the free tier. You also need the raw material your GPT will lean on: a clear description of its job, and any reference files (PDFs, docs, notes) you want it to draw from. Have those ready and the build goes quickly.

How to create a custom GPT, step by step

The flow is short, and you configure everything in plain language.

1. Open the GPT builder. In ChatGPT, go to your workspace and choose to create a new GPT. You will see two tabs: Create, a chat that builds the GPT for you by asking questions, and Configure, where you set everything by hand. The Configure tab gives you the most control.
2. Name it and describe it. Give the GPT a clear name and a one-line description of what it does. This is what shows up if you later publish it to the GPT Store.
3. Write the instructions. This is the core of the build. Spell out the GPT's role, its tone, the tasks it should handle, and the things it should refuse. Be specific. "Answer only from the uploaded handbook and say when you don't know" beats a vague persona line.
4. Upload knowledge files. Add the documents you want the GPT to reference, such as PDFs, spreadsheets, or text files. The model can pull from these when it answers. Keep them clean and current, because the GPT will treat them as source material.
5. Turn on abilities. Enable the capabilities the job needs: web browsing for current information, image generation, or code execution for data tasks. Only switch on what you actually need, since each ability widens what the GPT can do and where it can go wrong.
6. Test, then save and share. Use the preview pane to try real questions and watch how it answers. When it holds up, save it and choose its visibility: private (just you), link-only (anyone with the URL), or public in the GPT Store, where other ChatGPT users can find and run it.

For a task like "answer questions from this manual" or "draft posts in this format," that is often all you need, and you can be done in an afternoon.

What you can build with it, and where it stops

The steps above are enough to ship a genuinely useful helper in an afternoon. Point one at a returns policy so every refund question gets the same wording. Load a quarter of past newsletters and have it draft new sections in that cadence. Hand a team a single GPT that answers from an onboarding doc instead of pinging a channel. For narrow, repeatable jobs like these, a Custom GPT earns its place. (If you want the underlying concept rather than the build, see what a Custom GPT is.)

Where it stops is the moment the GPT has to stand in for you in public. Three limits surface fast once it does.

It won't show its work. A Custom GPT replies in the model's smooth voice but rarely names the source a claim came from. The files you upload are treated as one ingredient among many, so the model can quietly fold them into its own training and hand back a confident, unsourced answer. Language models are well documented as producers of fluent but false statements, and tying each reply to the exact passage it rests on drives that error rate down without ever reaching zero. Answers that carry a checkable citation are the ones a reader can trust in your name, and that is precisely the discipline a stock Custom GPT skips.

You are renting the address. Everything about your GPT lives inside ChatGPT. A visitor needs a ChatGPT account to reach it, it surfaces on OpenAI's terms, and those terms are OpenAI's to rewrite. Perfectly fine for an internal tool. A standing risk when the product itself carries your name.

The pricing isn't yours. Publishing to the GPT Store is possible, but the ranking, the terms, and any payout stay on OpenAI's side of the table. You don't name your own price or hold the billing relationship with the people using it.

When an owned AI twin is the better fit

When the thing carrying your name has to be dependable and answerable for what it says, a Custom GPT is the wrong container. An owned AI twin, a grounded, in-voice version of you built only from what you have published, is shaped for exactly that job. The differences line up point by point.

• Sources you can open. Every reply comes from your own library with the passage behind it on display, so a reader can hold the answer up against what you actually wrote. Anchoring to real sources lowers the invented-answer rate without erasing it, which is why at Twinsona the citation and the guardrail are the thing you are buying, not trim added after the model ships.
• Your register, not the assistant's. Ask a Custom GPT something pointed and you get ChatGPT's tidy, anonymous tone; ask a twin and it answers in the wording, examples, and positions already on your record, because your published work is the entire training set.
• Ownership. A twin keeps the audience, the price, and the use of your likeness on your side, each of them moving only when you say so. That is command over how it runs, not a downloadable model file you self-host.
• Pricing you set. Creators already put a price on their own AI: Tony Robbins and Matthew Hussey both sell creator AI at $39 a month from their own sites, and the platforms match the pattern, with Delphi listing an entry Builder plan near $79 a month. Twinsona lets you set what twin access costs, invoice your audience straight, and keep the proceeds, instead of surrendering the terms and the cut to a store.

Ready to build? Take the six steps above for a Custom GPT, or start the create an AI clone overview if an owned twin is the shape you need.

FAQ

How do I create a Custom GPT? Open ChatGPT on a paid plan, launch the GPT builder, then use the Configure tab to set a name, instructions, knowledge files, and abilities. Test it in the preview pane and save it as private, link-only, or public in the GPT Store. No coding is required.

Do I need to pay to build a Custom GPT? Yes. The GPT builder is only available on a paid ChatGPT plan. There is no separate per-GPT fee to create one, and publishing to the GPT Store is free, but any monetization runs on OpenAI's terms.

How long does it take to build a Custom GPT? Most builds are done in under an hour, and a simple single-task GPT can be ready in a few minutes. The time goes into writing clear instructions and preparing clean knowledge files, not into setup or code.

Why does my Custom GPT ignore its knowledge files? Uploaded files are one input the model can draw on, not a hard boundary, so it may blend them with its own training and answer without citing them. A tight instruction like 'answer only from the uploaded files and say when they do not cover it' helps, but grounded retrieval lowers made-up answers rather than removing them. An owned AI twin is built around that limit instead of against it.

When should I build an owned twin instead of a Custom GPT? Build a Custom GPT for a fast internal or single-task helper. Reach for an owned AI twin when the tool carries your name in public: a twin answers only from your content, cites what it says, speaks in your voice, and lets you set the price and bill your audience directly instead of publishing under a store's terms.

About the author

Ankur Shrestha is the founder of Twinsona, where he builds the grounding-and-guardrail layer that keeps a creator's AI twin faithful – answering only from the creator's own content, citing its sources, and never drifting from what they actually said. Before Twinsona, he built agentic AI automating insurance-carrier portals – high-stakes work where being wrong carries real consequences, the same accountability problem he now solves for creators.