To create a custom GPT, open the GPTs area in ChatGPT on the web, select Create, define one clear job, configure the instructions and tools, then test real prompts in Preview before you share it. A custom GPT is a configured version of ChatGPT. It is not a newly trained model, and it is not an app you can embed on another website.
OpenAI currently limits building and editing to paid ChatGPT accounts on the web. A managed workspace can add its own role and sharing restrictions. People you share it with can use it in ChatGPT where GPTs are supported, but the editor itself is a web feature.
Fast path: go to Explore GPTs and choose Create. Write the public-facing name, description, and instructions, then upload only the reference files you are allowed to use. Enable the minimum capabilities required, run a small test suite, and keep the first version private until it passes.
What a custom GPT is, and what it is not
OpenAI describes GPTs as versions of ChatGPT configured for a specific purpose. A GPT can combine instructions, conversation starters, uploaded knowledge, built-in capabilities, apps, or actions. The current GPTs in ChatGPT guide also makes two boundaries clear:
- GPTs run inside ChatGPT. Use the API if you need an assistant inside your own product or website.
- A GPT does not use saved memory, account-level custom instructions, or previous conversations. Each conversation starts fresh.
That second point changes how you should design one. If the assistant needs a customer number, preferred format, or project context, ask for it in the conversation or retrieve it through an approved connection. Last week's chat will not supply the missing context.
Define the job before you open the builder
A narrow job is easier to instruct and test than a broad persona. "Help with marketing" leaves too many decisions unstated. "Turn an approved product brief into three landing-page headlines and list every unsupported claim" gives you an input, an output, and a review point.
Write a one-page specification with five items:
- User: who will open the GPT and what they already know.
- Job: the single task it should complete.
- Required inputs: information it must request before working.
- Output: the exact format a satisfactory answer should follow.
- Boundaries: questions it should decline, escalate, or label as uncertain.
| Vague concept | Testable first version | Simple acceptance check |
|---|---|---|
| Research assistant | Summarize an uploaded policy and cite the relevant section | Every factual answer includes a file and section reference |
| Writing coach | Review a 500-word product update against a six-rule style guide | Feedback maps each issue to one named rule |
| Data analyst | Inspect one CSV template and return a fixed set of quality checks | The same test file produces all required checks in the agreed order |
Keep email, record changes, and spending outside the first version. A read-only workflow lets you understand the answers before actions add consequences. Add an action after its inputs, confirmation step, failure behavior, and owner are clear.
Open the GPT builder
The current OpenAI creation guide gives this path:
- Open Explore GPTs in the ChatGPT sidebar or visit chatgpt.com/gpts.
- Select Create.
- Describe the GPT conversationally, or configure its fields directly in the editor.
- Use Preview to test it before saving or sharing.
Interface labels can move. If an old tutorial points to a missing /gpts/mine page or tells you to find separate Create and Configure tabs, start again from Explore GPTs and follow the current editor.
Configure each field for a real user
Name, description, and conversation starters
The name and description appear before someone starts a chat. They should set expectations instead of carrying every keyword. The name identifies the job; the description says who the GPT is for, what input it expects, and what it returns.
For example, Support Policy Finder is clearer than Customer Success Genius. Its description could say: "Answers support agents' questions from the approved policy files and cites the section used." That wording also tells a reviewer what the GPT should not do.
Conversation starters are sample prompts shown when the GPT opens. Add a few that represent the main workflow, such as:
- "Which policy covers a refund after 30 days? Cite the section."
- "Compare the return rules for annual and monthly plans."
- "The files do not mention this case. What should I escalate?"
Use an icon that remains legible at a small size and that you have permission to publish. The icon helps recognition, but it cannot rescue an unclear job or unreliable answers.
Instructions
Instructions apply to every conversation. Put behavior, workflow, tone, and boundaries here. Put source material in Knowledge. OpenAI recommends explicit steps, positive and concrete directions, headings, and short examples when a definition or classification matters.
This compact template is a starting point:
Role
You help [user] complete [specific job].
Required input
Ask for [items] before starting. If an item is missing, ask one concise question.
Workflow
1. Check whether the request is in scope.
2. Use only [allowed sources or tools].
3. Complete [named checks] in this order.
4. Mark uncertainty and missing evidence.
Output
Return [format, sections, length, citations].
Boundaries
Do not invent facts, quotes, sources, or completed actions.
Escalate [high-impact or ambiguous cases] to [responsible person].Replace every bracketed item. Then add one good example and one unacceptable example if the task has a strict format. Avoid a page of overlapping rules. Contradictory instructions make failures harder to diagnose.
Knowledge files
Knowledge gives the GPT reference material. Upload the handbook, guide, catalog, or other source it should consult. Operating rules belong in Instructions, where they are visible and editable. The document being cited belongs in Knowledge.
Text-forward files with clear headings are easier to test than scans and complex page layouts. Remove obsolete drafts, duplicate versions, hidden comments, credentials, personal data, and material you are not allowed to upload. If the GPT must quote or cite a file, specify the expected citation format in Instructions and verify it in Preview.
Check the live limit before planning a large knowledge base. As reviewed on July 17, 2026, OpenAI's GPT editor guide says a GPT can have up to 20 files at 512 MB each, while the separate File Uploads FAQ still lists 10 files per GPT. The official pages do not agree. Treat the editor shown to your account as the operational limit and avoid designing a workflow that depends on filling it exactly.
Recommended model and capabilities
A recommended model is a suggestion for the user starting a conversation. It is not a permanent lock. Users may be able to switch models, and ChatGPT can substitute a similar model when the recommendation is unavailable. Test behavior again after a model change instead of naming a model as a guarantee.
Enable only the capabilities required by the job:
| Capability | Use it when | Test before sharing |
|---|---|---|
| Web search | The answer needs current public information | Sources are relevant, dated, and not substituted for your Knowledge files |
| Image generation | The task explicitly produces images | Prompts, brand constraints, and unsafe requests are handled as intended |
| Canvas | Users need to draft and revise longer structured content | The workflow still follows your required sections |
| Code Interpreter and Data Analysis | The GPT must calculate, inspect files, or create charts | Known inputs produce correct values and clear error messages |
Apps and actions
Apps connect to services available to the user. Actions connect to external APIs that you define. A GPT can use apps or actions, but not both at the same time.
An action needs API details, an authentication choice, and an OpenAPI schema. OpenAI's actions configuration guide covers no authentication, API keys, OAuth, schemas, and workspace domain restrictions. Keep secrets out of instructions and schema descriptions. Test every action in Preview, including missing authorization, rejected input, a slow response, and a failed upstream service.
Be explicit about data flow. Relevant parts of a user's input may be sent to a connected app or external API. OpenAI states that it does not control how those third parties use or store the data. Use only services you trust, disclose what leaves ChatGPT, and request the minimum data the action needs.
Test the GPT in Preview
A polished answer to one happy-path prompt is not a release test. Build a small set of prompts and save the expected result for each one. Re-run the same set after changing instructions, files, tools, or the recommended model.
- Core tasks: representative requests with complete inputs.
- Missing inputs: requests that should trigger a clarifying question.
- Unsupported questions: topics the GPT should decline or escalate.
- Source gaps: questions the uploaded files do not answer.
- Conflicting sources: two files that disagree or have different dates.
- Hostile text: a file or user message that tries to override the GPT's rules.
- Tool failures: unavailable web results, malformed files, or rejected actions.
Record pass, fail, and the reason. Fix the largest failure category, then run the full set again. Adding more tools is rarely the first fix; OpenAI recommends tightening instructions and examples before expanding the configuration.
OpenAI says GPT builders cannot view individual conversations users have with their GPTs. Collect feedback through a separate, privacy-conscious channel, and describe that channel without suggesting that private chats are visible to the builder.
Choose how to share or publish
Keep the first working version private while you test. OpenAI's current sharing and publishing guide lists options that vary by plan and workspace:
- Invite-only: only you on a personal account, or selected recipients in an eligible managed workspace.
- Workspace sharing: specific people, groups, link access, or the whole workspace when an administrator allows it.
- Anyone with the link: public link access when available.
- GPT Store: a public listing for an eligible GPT and account.
Managed workspaces can also offer Can chat, Can view settings, or Can edit permissions. Public link and Store access support chat use, not public viewing of the configuration.
Store publication may require a builder profile, a category, and policy review. A public website in the builder profile must use a verified domain. If the GPT uses actions, each public action needs a valid Privacy Policy URL. An app connection, workspace restriction, or policy check can also block public publication.
Publishing is not the end of testing. Edits save to a draft; choose Update to apply them to the live GPT. Version history can help you inspect or restore an earlier version, although an action's authentication may need to be configured again after a restore.
Pre-publish checklist
- The name and description state one job without promising perfect results.
- Conversation starters demonstrate the main tasks.
- Instructions define inputs, steps, output, uncertainty, and escalation.
- Knowledge files are current, authorized, readable, and free of secrets.
- Only necessary capabilities are enabled.
- Apps or actions have a clear data-flow explanation.
- Every action has failure tests and a confirmation step where consequences matter.
- The saved test set passes after the latest change.
- The sharing level and permissions match the intended audience.
- Public actions have a valid privacy policy and the builder profile is ready.
Questions that change the build
Can free users create a custom GPT?
OpenAI currently says creating and editing require a paid ChatGPT subscription. Signed-in users can use GPTs they have access to. Workspace permissions may add further limits.
Can I embed a custom GPT on my website?
No. Custom GPTs are designed to run inside ChatGPT. Use the OpenAI API to build an assistant into a website or product.
Does a custom GPT remember users between chats?
Not through saved memory, custom instructions, or previous GPT conversations. Each conversation starts fresh, so request the necessary context or retrieve it through an approved connection.
Can GPT builders read user conversations?
No. OpenAI says builders cannot view individual conversations people have with their GPTs. Inputs sent to an app or external API are subject to that third party's practices, which is a separate privacy consideration.
Should information go in Knowledge or Instructions?
Put reference material in Knowledge. Put rules, workflow, tone, output format, and citation requirements in Instructions. Test both together in Preview.
Should I use apps or actions?
Use apps for supported services connected by the user. Use actions when you define an external API with an OpenAPI schema and choose the appropriate authentication mode. The same GPT cannot use both at once.
