How to Create a Custom GPT for Your Business Step by Step

Don't build a new one: OpenAI is retiring custom GPTs, which stop running on 11 December 2026, and it has been switching off the option to create new ones. Build the same assistant as a ChatGPT Project (instructions plus files, shareable), or as a skill or plugin on a Business workspace. If you already have a GPT, migrate it before December.

The need behind this title hasn't gone away. Owners want an assistant that already knows their products, prices and tone, so staff don't re-explain the business in every chat. That's still easy to build; it just lives somewhere else now. A GPT's instructions become project instructions or a skill, its knowledge files become project files or reference files, and its conversation starters become saved prompts. Only custom actions, where a GPT called another system, need real rebuilding.

Follow me on Instagram@sagnikteaches

What's happening to custom GPTs, in dates

The details come from OpenAI's custom GPT retirement and migration FAQ and its migration guide:

Connect on LinkedInSagnik Bhattacharya

• 11 December 2026: custom GPTs stop running on all ChatGPT plans. Enterprise workspaces with an approved deferral get until 11 February 2027.
• Until then: existing GPTs keep working, subject to their current sharing settings. OpenAI says dates can change and each account gets its own notices.
• Creation: the ability to create new GPTs is being withdrawn, with timing that varies by plan and workspace.
• Migration: a published GPT can be turned into a plugin. Its instructions become a skill, its knowledge files become reference files, and connected apps come across.
• What doesn't transfer: custom actions (anything calling an outside system through an API), the selected model, and possibly conversation starters and previous chats.
• After migrating: the original GPT becomes read-only but still runs until retirement. The new plugin starts private; old sharing and links don't carry over.

So there are two jobs. If you're starting from nothing, choose a replacement and build it (the next three sections). If you already have GPTs, migrate or rebuild them in time (the section after that).

Subscribe on YouTube@codingliquids

Which replacement fits what you wanted a GPT for

For most small businesses the ChatGPT Project is the right first move: it's on every plan, it's shareable, and it covers what most business GPTs were actually used for. The last row matters too; our tutorial on whether a custom GPT should answer your customers explains why customer-facing assistants need different tools.

Option 1, step by step: build it as a ChatGPT Project

An illustrative garden centre wants what it would once have built as a "Plant helper" GPT: staff ask about care, hardiness and companion planting, and answers come from the centre's own care sheets and stock list, in the centre's voice.

1. Create the project. In ChatGPT's sidebar, select New project and name it "Plant helper".
2. Write the project instructions. Open the project's settings and add instructions. These apply to every chat in the project, the way a GPT's instructions did.
3. Add files. Upload the care sheets, the current stock list and a style guide. Limits per project: 5 files on Free, 25 on Go and Plus, 40 on Pro, Business, Edu and Enterprise.
4. Set memory to project-only in Project settings, so the project draws on its own chats rather than everything else the account has discussed.
5. Test with ten real questions staff asked last month, and fix the instructions until the answers are right.
6. Share it. Shared projects are available on all plans. Collaborator limits are 5 on Free, 10 on Go and Plus and 100 on Pro, and you can give people Edit or Chat access. Choose "only those invited" rather than "anyone with a link" for anything internal.

The instructions the garden centre used:

You help garden centre staff answer customers' plant questions.
Use the care sheets and stock list in this project first. If they don't cover the
question, say "Not in our sheets" and give a short general answer labelled GENERAL.
Rules:
- Say whether we have the plant in stock only if the stock list shows it.
- For anything about plants and pets or children, quote the care sheet's toxicity
  note exactly. If there isn't one, say "Check before recommending".
- British spelling, friendly, under 100 words, no Latin names unless asked.
End every answer with: Source: [file name] or GENERAL.

An illustrative test, and what it caught. Question: "Customer wants something evergreen for a shady fence, not too big." First answer: three shrubs, two from the stock list and one that sounded right but wasn't stocked, with no source line. The fix was adding "only recommend stocked plants unless asked" and putting the source rule last, where it was followed more reliably. Second answer: two stocked shrubs, height and spread from the care sheets, source line present.

For more on running separate projects for different clients or areas of work, see using ChatGPT Projects to keep client work separate.

Option 2: a skill or plugin on a Business workspace

If your team is on ChatGPT Business or Enterprise, skills and plugins are OpenAI's direct replacement for GPTs. The difference from a project is how people use them: a skill is called up by name (with an @ mention) from any chat, rather than living in one project.

To create a skill, go to Skills, select Create, then either Create with chat (ChatGPT asks follow-up questions and builds it) or Create with editor (you write it). A good skill does one focused task with clear inputs and outputs, such as "turn a grower's availability sheet into point-of-sale labels".

To create a plugin, which can bundle skills with connected apps, OpenAI's guide describes five steps:

1. Mention Plugin Creator in a chat.
2. Describe the workflow: its purpose, inputs, outputs and rules.
3. Add context: templates, examples or reference files.
4. Review the name, description and instructions it drafts.
5. Test it on familiar tasks before sharing.

You need the "Use plugins" permission and Plugin Creator enabled in your workspace; admins control who can share plugins and publish them to the workspace. Plugins start private, so share deliberately. Anyone you share with must install it unless an admin pre-installs it, and they need access to any connected apps it uses. On personal plans, availability of skills and plugin creation has been rolling out gradually; if you don't see Skills in your sidebar, use a Project.

Which to choose on a Business plan? A project for an area of knowledge people discuss together; a skill for a task done the same way every time. The garden centre ended up with both: the "Plant helper" project for care questions and a "Label writer" skill for the weekly batch of new stock labels.

Turning GPT instructions into a skill that behaves

GPT instructions were often written as a personality ("You are a friendly assistant for...") with rules scattered through them. Skills work better written as a procedure with clear inputs and outputs. A before and after from the garden centre's label task:

BEFORE (GPT instructions)
You are Label Buddy, a cheerful helper for our garden centre! You know loads about
plants. When someone gives you plants, make nice labels. Keep it short. Use British
spelling. Don't make stuff up!

AFTER (skill)
Purpose: turn a grower's availability sheet into point-of-sale plant labels.
Input: rows pasted from the availability sheet (name, Latin name, pot size, price).
Steps:
1. For each row, write: common name; Latin name copied exactly; height and spread
   from the care sheet reference file; position (sun/shade); one line on why buy it.
2. If the care sheet has no entry, write [NO CARE SHEET] and skip the extra lines.
3. Price exactly as given, with pot size.
Output: one label per row, separated by a line, ready to paste into the label template.

The "after" version tells ChatGPT what comes in, what to do in order, what to do when data is missing and what comes out. That structure is what makes a skill reliable enough for someone else on the team to call up without supervision.

Option 3: a Claude Project or a Gemini Gem

A GPT's instructions and files work almost unchanged in the other two big assistants. A Claude Project holds instructions and knowledge files and can be shared across a Team or Enterprise organisation; our tutorial on setting up Claude Projects as a shared team assistant covers it. A Gemini Gem suits businesses on Google Workspace, especially for one fixed task, and building Gemini Gems for repeat tasks goes through it. For a side-by-side view, see custom GPT vs Claude Project vs Gemini Gem.

Migrating a GPT you already have, before 11 December

Start with an inventory. OpenAI's own advice is to record each GPT's creator, instructions, reference files, integrations and some familiar test prompts. A filled-in example:

GPT name:        Plant helper
Created by:      owner's work account
Used by:         6 staff (shared by link)
Instructions:    copied to "GPT backups" doc, 14 Oct
Knowledge files: 6 (care sheets x4, stock list, style guide) - stock list out of date
Actions:         none
Starters:        "Is this plant safe for dogs?", "What can go in a shady border?"
Test prompts:    10 real staff questions saved in the backups doc
Replacement:     ChatGPT Project (shared), same files, updated stock list
Deadline:        replacement live and staff switched by 1 Dec

Then, for each GPT:

1. Back up the instructions and conversation starters into a document you own. Starters may not carry over.
2. Migrate or rebuild. Where migration is available, go to My GPTs, find the GPT under Created by me, and select Migrate to plugin. Only the creator or a workspace admin can migrate, the GPT must be published, and plugins must be enabled in the workspace. If you'd rather use a Project, rebuild it by hand from your backup.
3. Test against your saved prompts, plus one difficult case. Check the right skill is picked up, reference files are used, and the output format still matches. The model may differ from the one you chose for the GPT, so behaviour can shift.
4. Rebuild any custom actions through a supported connected app or a custom MCP server. MCP (Model Context Protocol) is a standard way of connecting AI tools to other software; this step usually needs whoever built the original action.
5. Share the replacement with the same people. Sharing doesn't carry over, and bulk migrations by admins don't copy sharing settings either.
6. Tell users and set a switch-over date well before 11 December, so problems surface while the old GPT still works.

If Migrate to plugin doesn't appear, run through the usual causes: the migration experience hasn't reached your account yet (OpenAI is rolling it out by plan and workspace); you're not the GPT's creator or a workspace admin; the GPT was never published (publishing it for migration doesn't mean sharing it publicly); or plugins are switched off in your workspace settings. If it still isn't available a few weeks before the deadline, rebuild by hand as a Project from your backup rather than waiting.

A realistic timetable for a small business: inventory by the end of October, rebuild and test in November, switch staff over by 1 December, and leave the old GPTs running untouched until they stop.

A garden centre's two GPTs, moved in an afternoon

The illustrative garden centre had two GPTs on ChatGPT Business: "Plant helper" (shared with six staff) and "Newsletter writer" (used by the owner to draft the monthly email). Neither used custom actions.

Plant helper became a shared ChatGPT Project. Rebuilding it took about 40 minutes: copying the instructions, uploading the six files (with a current stock list replacing the stale one), and running the ten saved staff questions. Two answers changed: one improved because the stock list was now current, and one got longer, fixed by adding "under 100 words" to the instructions. Sharing it with six staff took five minutes.

Newsletter writer was migrated with Migrate to plugin, so its instructions became a skill. The migration took minutes; testing took half an hour, because the first draft came out more formal than before. Adding two past newsletters as reference files brought the tone back.

Total: under two hours of the owner's time, done in October, with six weeks of margin before the old GPTs stopped. The figures are illustrative, but the proportions are typical: moving a GPT is quick; testing that it still behaves takes most of the time.

When the GPT had a custom action: a picture framer's order checker

Actions are the one part that needs real work, so here's how one illustrative case went. A picture framer had a GPT that staff used to answer "is my frame ready?" calls. It had a custom action calling a small script that read the order-tracking Google Sheet: staff typed an order number and the GPT replied with the status and collection date.

Migration moved the instructions into a skill, but the action didn't transfer, so the plugin could no longer see the Sheet. There were three ways forward, in order of effort:

1. Connect Google Drive as an app in the workspace, so the skill can read the order Sheet through a supported connection. This was the route the framer took, and it needed no code: an admin enabled the Google Drive app, each staff member authorised their own account, and the skill's instructions were changed to "look up the order number in the Orders sheet".
2. Rebuild the action as a custom MCP server, which suits systems with no ready-made connection, such as a bespoke order database. That's a developer's job.
3. Drop the lookup and have staff check the Sheet directly. Sometimes the honest answer is that the GPT was saving ten seconds per call and isn't worth rebuilding.

The test that mattered: five real order numbers, including one that didn't exist. The first version invented a status for the missing number. Adding "if the order number isn't in the sheet, say so and stop" fixed it. Any assistant that reads live data needs that rule, whatever it's built in.

Mistakes to avoid during the move

• Building a new GPT now. Even where the Create button still appears, anything built as a GPT stops working in December.
• Assuming sharing carries over. Staff following an old GPT link won't land in the new plugin or project. Send new links and check each person can open them.
• Skipping the test. "It migrated" isn't "it works". Run the saved prompts every time.
• Forgetting actions. A GPT that looked up stock or booked appointments will fail quietly in the plugin until the integration is rebuilt.
• Losing the prompts people liked. Save conversation starters into a shared prompt library.
• Relying on other people's GPTs. Public GPTs made by others retire too, and you can't migrate them.

If you're deciding plan levels at the same time, ChatGPT Plus vs ChatGPT Business covers which plan a team needs for sharing and admin controls.

Custom GPT retirement: what owners ask

Will my existing custom GPT keep working until December?

Yes. OpenAI says existing GPTs stay usable until the retirement date, subject to their existing access settings. After migration the original becomes read-only but still runs until retirement. On 11 December 2026 custom GPTs stop running; approved Enterprise deferrals run to 11 February 2027. Plan to have the replacement tested well before then.

What about public GPTs made by other people that my team uses?

They retire on the same timetable, and you can't migrate someone else's GPT. List any third-party GPTs your staff rely on, check whether the maker has announced a plugin replacement, and if not, rebuild the useful part as your own Project or skill using the prompts your team typically gives it.

Do I need a Business plan to replace a GPT?

Not for the simplest replacement. ChatGPT Projects, with their own instructions and files, are available on every plan, including Free with lower limits, and can be shared. Skills and plugins built for a whole team are designed around Business and Enterprise workspaces. A sole trader on Plus can usually rebuild a GPT as a Project in under an hour.

Can I just copy my GPT's instructions into Claude or Gemini instead?

Yes, and it's a reasonable time to review the choice. A GPT's instructions paste straight into a Claude Project or a Gemini Gem, and knowledge files upload the same way. Expect to adjust the wording after testing, because each model reads instructions slightly differently. The retirement is also a reminder to keep your instructions in a document you own.

Further reads

• How to Train ChatGPT on Your Business Information Safely — How to give ChatGPT your business context safely, beyond one assistant.
• RAG vs Custom GPT vs Fine-Tuning: What Does Your Business Need? — When a bigger, retrieval-based setup is worth it instead.
• AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep instructions and files portable so the next change hurts less.
• What If Your AI Vendor Shuts Down? Checks Before You Commit — Supplier-risk checks the GPT retirement makes very concrete.
• How to Connect ChatGPT or Claude to Your Business Apps — Rebuild a GPT's actions through connected apps instead.
• 12 Custom GPT Examples for Small Businesses, With Setup Notes — Twelve assistants small businesses built as custom GPTs, rebuilt as Projects and Gems after the GPT retirement, each with instructions, files and sharing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center, Custom GPT retirement and migration FAQ; ChatGPT Learn, Moving your custom GPT workflows to plugins, Build plugins, and Skills and plugins; OpenAI Help Center, Projects in ChatGPT; OpenAI announcement on shared projects (all checked September 2026).