How to build a Custom GPT or plugin for your Business

A custom GPT is a version of ChatGPT set up once for one recurring job, so that nobody has to type the same long brief again. Many businesses built one in 2024 and 2025. In September 2026 OpenAI announced that it is retiring them, which changes how you should build the next one.

This guide covers both halves. It explains how a custom GPT is built, step by step, because the thinking behind it carries over to anything you build next. It then explains the plugin, the format OpenAI is moving everything to, and how to build one or move an old GPT across before the deadline. It is written for a small business, so it leaves out anything that needs a developer unless that is stated.

Key takeaways

• A custom GPT is a ChatGPT assistant with its own instructions and, if needed, its own documents, set up for one job.
• OpenAI has announced that custom GPTs are scheduled to stop running on 11 December 2026. Creating new ones is already closed on personal accounts.
• The replacement is the plugin: a package of skills (reusable instructions), reference files and connected apps that works inside any chat.
• The hard part is the same in both formats: a clear job, good instructions, clean documents and a person who tests the result.
• Do not put anything into an assistant that you would not paste into ChatGPT directly. Builders cannot read individual chats, but the files and instructions you add are not secret from the people who use it.
• Build one only for a job that is frequent, rule-bound and checkable. A saved prompt is often enough.

What is a Custom GPT?

A custom GPT is a version of ChatGPT that has been briefed in advance. You write down the job, the tone, the format and the rules once. You can attach documents it should answer from. Whoever opens it starts with an assistant that already knows what it is for, instead of a blank chat box.

It is not a new model and it is not trained on your data in the technical sense. It is the same ChatGPT with a standing brief wrapped around it. That is why it is quick to make and why it is only as good as the brief you write. The site’s glossary page on the custom GPT sets out the idea in more detail, and the ChatGPT Enterprise course builds five worked examples: a prompt improver, an HR assistant that answers from the staff handbook, an expense categoriser, a communication coach and a simulator for practising difficult conversations.

Three older features get confused with it. The first, custom instructions, are standing rules that apply to everything you do in ChatGPT, not to one job. A prompt library is a set of saved prompts that you pick from and paste. Memory is something the tool builds up by itself from your chats. A custom GPT sits apart from all three: you build it on purpose, for one task, and you open it when that task comes up.

Are Custom GPTs being retired?

Yes. OpenAI’s ChatGPT release notes for 11 September 2026 say the company plans to retire custom GPTs across ChatGPT plans and move people to plugins. OpenAI’s migration FAQ and the guides that summarise it give this timeline:

• Personal accounts (Free, Go, Plus and Pro) can no longer create or publish new GPTs. Existing ones can still be used, and edited by people on an eligible subscription.
• Enterprise workspaces stop creating new custom GPTs ahead of the main deadline. The sources give different dates for this (late September or 26 October), so read the notice inside your own workspace.
• 11 December 2026 is the scheduled date when custom GPTs and their pages become inaccessible.
• 11 February 2027 is the later deadline for Enterprise workspaces that have an approved deferral. Nobody else gets the extension.

OpenAI says these dates may change, and that Free, Plus and Pro have no firm dates beyond the creation block, so treat the main date as a target to plan around, not a promise. What matters for a business owner is the direction. If you or your team rely on a custom GPT today, it needs a plan. If you were about to build one, build the plugin version.

What is a ChatGPT plugin, and how is it Different from a Custom GPT?

A plugin bundles reusable capabilities into one installable package. According to OpenAI’s developer documentation, a plugin is built to combine skills, MCP servers and optional interface elements. In plain terms it can contain three kinds of thing:

• Skills. Reusable instructions for a repeatable workflow. This is where the old custom GPT instructions go.
• Reference files. The documents the work should be based on, the equivalent of knowledge files.
• Connected apps. Links to outside services such as email or file storage, increasingly built on the Model Context Protocol, the open standard for connecting AI tools to other systems.

The difference that matters day to day is where it lives. A custom GPT was a separate chat. You opened it, and everything in that conversation followed its brief. A plugin works inside any chat. You can call it with an @ mention, or ChatGPT may decide to use it when your request seems to match. OpenAI is clear that the plugin will not necessarily run on every request. That is a real change: with a custom GPT you knew the brief was always on, and with a plugin you name it when it matters and check that it was used.

If you want a deeper explanation of the connection layer, MCP and an API compares the two ways of linking an AI tool to your systems.

What can you use a Custom GPT or plugin for in a Business?

Use one for a task that comes round often, follows rules that can be written down, and produces a result a person can check. The pattern is more reliable than any list, but a few examples show it:

• Answering staff questions from the handbook. The questions repeat, the answers sit in a document you own, and a wrong answer matters, so the assistant should quote the passage it used.
• Drafting a standard reply. A support or sales email that follows your tone and always includes the same three points.
• Sorting and labelling. Expense lines, enquiries or invoices put into your categories, with anything unclear flagged for a person.
• Reviewing writing against your standard. A style guide in the assistant, and a rule that it lists problems before it rewrites anything.
• Practising a conversation. A simulator that plays a difficult customer or candidate and stays in the role.

The jobs that suit it badly are one-off questions, tasks whose rules change every time and anything where a wrong answer reaches a customer with nobody looking. If you are not sure where to begin, the guide to the first AI tasks a small business should hand over sorts jobs by risk. The wider question of whether to use a ready-made tool or build something custom is covered in off-the-shelf tool or a custom workflow.

How do you build a Custom GPT step by step?

The mechanics are simple. The thinking is where the effort goes. Here is the sequence, in the order that saves rework. Use it for the plugin too, because the same pieces are involved.

Step 1: write the job in one sentence. “This assistant answers staff questions about leave and expenses using only the staff handbook.” If you cannot say it in a sentence, the job is too broad and the assistant will be vague. Narrow it until you can.

Step 2: collect three real examples. Find three real requests the assistant will get and the answer you would want to each. These become your test set later, and writing them first shows you what the instructions must say.

Step 3: write the instructions. The next section covers how, because it is the part that decides quality.

Step 4: add only the documents the job needs. Reference material goes in the files. Behaviour rules go in the instructions. Keep them separate.

Step 5: choose the abilities. The builder offers options such as web search, image generation and data analysis. Switch on what the job needs and nothing else. A handbook assistant should not browse the web and then mix outside advice into your policy.

Step 6: test it with your three examples and ten more. Read every answer. The testing section below explains what to look for.

Step 7: decide who can use it and who owns it. Share it with the smallest group that needs it, and write down a named owner who updates the files and the instructions. An assistant nobody owns goes stale, which is the failure the guide to what happens when nobody owns the AI describes.

In the builder itself you can describe the assistant in a conversation, and it drafts the instructions for you, or you can write the instructions and settings by hand in the configuration view. Either way, read what it produced. A draft written by a builder is a starting point, not a finished brief.

How do you write Good Instructions for a Custom GPT?

Instructions are the brief the assistant carries into every chat, and most disappointing assistants fail here. The principles are the same ones that apply when you brief an AI tool like a new hire: give it context, a clear job, the format you want and the limits. Most weak answers trace back to a thin AI brief, not to the model. A useful structure has six parts:

1. Role and job. One or two sentences. “You answer staff questions about leave, expenses and working hours for a small company.”
2. Source of truth. Say which material it must use. “Answer only from the attached handbook. If the answer is not there, say you do not know and suggest asking the office manager.”
3. Tone and format. “Reply in plain English in three short paragraphs or fewer. Start with the answer, then the rule behind it.”
4. Evidence. “Quote the relevant sentence from the handbook after each answer.”
5. Limits. “Do not give legal advice. Do not discuss pay for named individuals. Hand any complaint or grievance to a person.”
6. Flags. “If a request looks like a complaint, a safety problem or a question about someone’s personal data, stop and say it needs a person.”

Two habits make a big difference. First, say what to do when the assistant does not know. Models are shaped to produce an answer, and an explicit permission to say “I do not have that” is one of the simplest ways to cut invented answers. Second, write rules as instructions, not wishes. “Be accurate” does nothing. “Quote the sentence you used” can be checked.

Keep the instructions short enough to read in one go. A wall of rules is harder for the model to follow and harder for you to maintain. If the instructions run past a page, the job is probably two jobs.

How do you add your own Documents to a Custom GPT?

Documents turn a general assistant into one that is grounded in your own material. Instead of answering from general knowledge, it answers from the text you supplied. That is the whole point of a handbook assistant, and it is the same idea that lies behind retrieval-augmented generation.

OpenAI’s support page on creating and editing GPTs gives the current limit as up to 20 files per GPT, each up to 512 MB. Older pages quoted other numbers, so check the figure in your own editor. A few rules make files work better:

• Use short, text-based files. Complex layouts, scanned pages and tables inside images are harder for the assistant to read. A clean text document or a simple PDF is better than a designed brochure.
• Keep files current. A grounded assistant is only as right as the documents behind it. An out-of-date price list produces confident, well-sourced, wrong answers.
• Put rules in the instructions, not the files. OpenAI’s guidance is that knowledge files are for reference material and the instructions are for behaviour.
• Ask for citations. If you want the assistant to point to its source, say so in the instructions, and say how: the file name, the heading or a short quote.
• Mind the context window. An assistant can only hold so much text in view at once. If you upload a very large document, it may work from parts of it, so split long material into sensible sections.

Before uploading, trim the file. A spreadsheet carries every row and column, hidden ones too, and a document can hold comments and tracked changes. The guide to what is safe to paste covers how to clean a file and what to leave out altogether.

How do you build a plugin in ChatGPT?

A plugin is built the same way you would build a custom GPT: by describing a job, supplying instructions and adding files. The difference is the tooling. Based on OpenAI’s support pages and the guides written about them, the route is this:

1. Open the plugin directory in ChatGPT and install Plugin Creator.
2. Start a chat with @plugin-creator and describe the workflow you want, in the same one-sentence form as step 1 above.
3. Choose what the plugin should contain: skills (your instructions), apps (connections to other services) or both. Add your reference files.
4. Test it with your real examples, then check the permissions before you share it.

Who can create plugins, and who can create skills, depends on your plan, your region and your workspace settings, and the published information is not consistent. One guide reports that personal skills are limited to Business, Enterprise, Edu and Healthcare workspaces, while OpenAI’s own page only says that installing and using a plugin depends on plan, workspace, role and region. Check what your account allows before you plan around it.

Developers can go further. OpenAI’s developer documentation describes building a focused set of tools on an MCP server, writing skills that use them, packaging both into a plugin and testing it in ChatGPT before submitting it to the directory. That route suits a business with its own software, and it is not needed for an assistant that only answers from documents and follows written rules.

One caution applies to every connected app. Only connect servers and apps from operators you trust. A file or web page that an assistant reads can carry hidden instructions, and a connected app widens what those instructions can reach. The article on prompt injection explains the risk, and it is a good reason to keep a plugin’s permissions narrow.

How do you move an Existing Custom GPT to a plugin?

If you already have a custom GPT, do not wait for the last week. OpenAI offers a migration route, and the safest approach is to back up first and migrate second. Guides written from OpenAI’s FAQ describe this order:

1. Back up. Copy the instructions into a document, download the knowledge files, note any custom actions, and save five to ten requests with answers you were happy with. There is no one-click export of a whole GPT.
2. Finish and publish your edits. Only the latest published version migrates. Drafts do not.
3. Migrate. Open My GPTs and choose Migrate to plugin. Only the creator can do this, or an admin in an Enterprise workspace.
4. Test. Run your saved requests through the plugin and compare. OpenAI notes that a migrated plugin may respond differently.
5. Rebuild actions. Custom actions do not transfer. If your GPT called an outside system, look in the plugin directory for a ready-made app first, and plan for a developer if you need your own.
6. Share and switch. The migrated plugin starts private. Share it, have one person try it, then move everyone over.

What carries across, according to OpenAI’s materials: instructions become a skill, knowledge files become reference files and connected apps become apps in the plugin. What does not: custom actions, past conversations, sharing settings and the model you selected. The original GPT stays usable until retirement but becomes read-only.

If you only use other people’s GPTs, your access does not automatically follow. Ask the creator whether they are migrating and whether they will share the replacement with you.

How do you test a Custom GPT before your team uses it?

Testing is where a decent assistant becomes a reliable one, and it is the step most people skip. Run it against your three real examples and at least ten more, taken from actual work, and judge each answer against a simple checklist.

• Is it right? Check every fact against the source document. An assistant can quote a policy that does not say what it claims.
• Did it stay inside the documents? Ask a question the files do not answer. A good assistant says so. A bad one invents something plausible, which is the way AI tools hallucinate.
• Did it follow the format and tone? If you asked for three short paragraphs and a quote, did you get them?
• Does it hold the line? Try to push it past its limits. Ask for legal advice, ask about a named colleague’s pay, ask it to ignore its rules. Note what it does.
• Is it consistent? Ask the same question three times. Different answers each time are a sign of guessing.

Then fix the instructions, not the answers, and run the set again. When it passes, write down the date and what you tested. Keep a human in the loop for anything that commits the business: the checks in the guide on how to reduce AI hallucinations apply to a custom assistant exactly as they do to a plain chat.

Is it Safe to put Business Data into a Custom GPT?

It is as safe as the account it runs in and the files you load. The assistant gives you no extra protection, and building it does not make the data inside it private.

On OpenAI’s own pages, builders cannot view individual conversations that users have with their GPTs, and work workspaces (Business, Enterprise and Edu) do not use workspace data to train models by default. Personal plans can use conversations for training unless the user opts out in the data controls. Those are useful facts, and they are not the whole picture:

• The instructions and files are visible to the assistant’s users in practice. People can often coax an assistant into repeating its instructions or quoting its files. Never put in anything you would not show every user.
• Outside services get part of what people type. If the assistant calls an outside service, relevant parts of the input may be sent there, and OpenAI does not control how that service stores it.
• Account type matters more than brand. The same assistant has different protection on a personal plan and a managed workspace.

Sort what you want to load before you load it. The usual tiers from data classification apply: public material is fine, internal material needs a managed account, and personal data, client confidential files and anything regulated stay out of general AI tools however neatly the assistant is set up. Under the GDPR, personal data in a file is still personal data, and a customer or employee has not agreed to it being in your assistant.

Finally, write the rules down. A one-page AI use policy says which accounts the team may use and what may go in them. Without one, a well-meant assistant built on a personal account is how shadow AI starts.

Do you need a Custom GPT, or is a saved Prompt enough?

Often a saved prompt is enough, and it is the simpler answer. Use the lightest option that does the job:

• A saved prompt in a shared prompt library, when the job needs one good prompt and a person to choose it.
• Custom instructions, when the rule is true of all your work, such as spelling, tone and your role.
• A project with reference files, when the work is ongoing and the material sits together. The comparison of ChatGPT Projects with the equivalents in other tools shows how they differ.
• A custom GPT or plugin, when a job is frequent and rule-bound enough that holding the whole set-up in one assistant is worth the upkeep.
• An AI agent, when the work needs several steps with tools and little supervision. An assistant answers and drafts. An agent acts, and it needs much stricter checks.

A useful test is to count. If the job comes up fewer than a couple of times a month, a saved prompt wins. If several people do it weekly and the instructions never change, build the assistant.

What to do next

Pick one job that comes up every week and write it as a single sentence. Collect three real examples with the answers you want. Write the instructions in the six-part form above, add only the documents the job needs, and test it on thirteen real requests. Build it as a plugin if your account lets you, and if you already have a custom GPT, back it up and migrate it well before 11 December 2026. Name an owner, and put the review date in the diary.

If you would like to work through this with an expert, the ChatGPT Enterprise course covers custom GPTs with worked examples, and a discovery call is a quick way to find out where an assistant fits in your processes.

Sources

• ChatGPT release notes (OpenAI): the 11 September 2026 note announcing the plan to retire custom GPTs across ChatGPT plans and move people to plugins.
• Custom GPT retirement and migration FAQ (OpenAI): the migration route, the Enterprise timeline, what carries over to plugins and what does not, with the note that dates may change.
• Creating and editing GPTs (OpenAI): the current guidance on building a GPT, including the limit of 20 knowledge files of up to 512 MB each, and the advice to keep reference material in files and behaviour in instructions.
• GPTs data privacy FAQ (OpenAI): the statement that builders cannot view individual conversations that users have with their GPTs, and the note on third-party services.
• Plugin development for ChatGPT and Codex (OpenAI developers): the definition of a plugin as skills, MCP servers and optional interface elements, and the build, package, test and submit sequence.
• OpenAI Is Retiring Custom GPTs: What You Need to Know (SVA Consulting): an adviser’s summary of the 11 December 2026 date, what moves to a plugin, and suggested alternatives by purpose.
• Custom GPTs Are Retiring: Dates, Migration and What to Do (Mimic): a step-by-step migration checklist, the full timeline and the advice to back up instructions and files before migrating.