4 min read
Custom GPTs: configure a focused assistant and check who can create one
The idea in 30 seconds
A custom GPT is a conversational assistant configured for a particular task using saved instructions, optional reference files and selected tools. Its setup guides its responses, without guaranteeing that it follows every instruction or gets every fact right.
A colleague asks how to book the meeting room, and you explain the same policy again. A custom GPT in ChatGPT can reuse written instructions and a policy document when answering that question. Before building one, though, check access: the creation rules have changed.
As checked on October 14, 2026, personal Free, Go, Plus and Pro accounts cannot create or publish new GPTs. Business, Enterprise and Edu workspaces can still allow creation through their permissions. Existing GPTs remain usable and editable where permitted. Creation and editing happen on the web; mobile apps support using GPTs.
OpenAI schedules custom GPT retirement for December 11, 2026, with different dates possible in account notices, including approved Enterprise deferrals. It recommends migration to plugins, which package reusable instructions, files and connected apps. Check your notice before starting a workflow you intend to keep.
Separate the rules from the reference material
If your workspace permits creation, describe the task to the builder or fill in the configuration fields yourself. Instructions define the job, response format and handling of missing information. Uploaded knowledge files supply reference material. A room-booking policy belongs in a file; the rule to quote its relevant section belongs in instructions.
Use an approved, clearly dated document with readable text. A saved policy is not a live booking calendar: without a suitable connection, it cannot establish whether the room is currently free. The saved configuration also differs from memory. New GPT conversations do not recall previous chats, so provide details needed for the current request.
Compare two hypothetical responses
The following prompts and responses are invented illustrations, not a product test. Imagine that the supplied policy does not mention weekend bookings.
A request that leaves the source unclear
Help staff book the meeting room. Can we reserve it on Saturday?This reply invents permission and a procedure. It sounds useful, but the supplied policy does not support either claim.
A request with a source and a missing-information rule
Use only the attached room-booking policy for policy claims. Quote the section supporting each answer. If the policy does not cover a question, identify the missing information. Can we reserve it on Saturday?The second instruction gives you something to check: a source passage or an explicit gap. It does not guarantee accurate answers; a quoted section still needs comparison with the original.
A reusable instruction template
Help [AUDIENCE] with [TASK]. Use [DOCUMENT] as the source for policy facts. Produce [FORMAT]. Cite the relevant section. If the document does not answer the question, state what is missing. Ask for [REQUIRED_DETAILS] before drafting the answer.Replace every bracketed field before saving. In Preview, try a question covered by the policy, one absent from it and one that misstates a rule. Record the answer and its cited passage so you can compare after changing the instructions.
Three short correction requests
- Show the exact passage supporting that policy claim.
- Separate facts in the document from your suggestions.
- Identify which information is missing before answering.
Check access before sharing
Use only material approved for the intended audience. Turning off Code Interpreter is not a confidentiality guarantee: the assistant can still include document content in an answer. Review the sharing options your workspace permits, and distinguish permission to chat from permission to view settings or edit the GPT.
Builders cannot view individual user conversations through the GPT builder. External apps or actions can nevertheless receive relevant input. Review those services and their permissions separately; a configured assistant is not an automatic website integration.
Success means the assistant handles the room-policy questions and gaps you actually check, with references you can verify. You can rehearse a request aloud in the mobile app's voice mode where available, including limited free access; this does not create or validate a GPT. Before retirement, retest the replacement plugin and its sharing permissions rather than assuming migration preserves them.