How to Migrate a Custom GPT to a Skill in ChatGPT or Claude

Anyone who's moved from one house to another knows the junk drawer. Everything else in the kitchen packs itself. The plates get wrapped, the pans go in a box marked PANS, and the table goes on the truck. Then you open the drawer by the back door and find a spare key, two takeout menus, the manual for a dishwasher you sold years ago, and a handwritten note about which breaker trips when the microwave and the toaster run at once. None of it belongs together. It only lived together because the drawer was there.

A move is the one time you're forced to decide what each of those things actually is. The key goes on the hook by the door. The menus go on the fridge. The manual goes in the recycling, and the breaker note goes somewhere you'll find it in the dark.

A custom GPT is that drawer. It holds instructions for doing a job, facts about your business that never change, a folder of templates, sometimes a few client files, and sometimes a connection to another system, all in one place because the GPT builder gave you one box to put them in. OpenAI is retiring custom GPTs on December 11, which means every one of those drawers is about to be moved. That makes this the best chance you'll get to sort one properly.

The move OpenAI offers, and what it leaves out

OpenAI's own path is a button. From My GPTs, you choose Migrate to plugin. The tool turns your GPT's instructions into a skill, copies its knowledge files across as reference files, and carries any connected apps. Custom actions, the model setting, your conversation history and your sharing settings stay behind. The dates and plan details are in OpenAI Is Retiring Custom GPTs. Here's the Official Record.

That button is worth using, and it answers a narrower question than the one you actually have. It moves the drawer as a unit. Everything your instructions said lands in one skill, whether it was a procedure, a standing fact about your agency or a tone rule. It also only exists in ChatGPT, and only where skills are switched on: OpenAI lists skills for Business, Enterprise, Healthcare and Edu workspaces, and says the migration button appears where it's available for your account. Personal plans aren't on the skills list. If you're on a personal plan, or you're thinking about moving some of this work to Claude, you're sorting by hand anyway.

Every GPT is holding three different things

Almost everything inside a GPT falls into one of three kinds, and each has its own new home.

A procedure. How to do one specific job: write a renewal summary, prep a podcast guest, turn meeting notes into a follow-up email. It has steps, a format, and a point where it's done. That's a skill.

Standing context. Things that are true whatever job you're doing. You work for Smith Agency, these are the carriers you're appointed with, and nothing in writing ever promises coverage. That belongs in a Project, in Claude or in ChatGPT, or in your custom instructions if it's really about you rather than the business.

Outside access. Anything the GPT reached out and touched, like a custom action that pulled policy data from your management system. That becomes a connector.

The reason to sort instead of shoveling everything into one skill comes down to when each thing gets read. A skill loads when the job calls for it, based on its description. Standing context has to be there every time. Put your agency's name and tone rules inside a skill and they only show up when that skill happens to fire. Put a twelve-step renewal procedure in your Project instructions and it sits in every conversation, including the ones that have nothing to do with renewals.

Client data is the hard line. A skill is built to be reused and shared, so client or case files go in a Project's files, where they stay with the work, and never inside a skill.

Here's what that looks like for an agency. Say your team built a renewal review GPT. Its instructions explain how to compare the expiring policy to the renewal offer and write a one-page summary for the client, which is the skill. They also say the agency's name, which carriers you write with and that the tone should be plain and never salesy, which is Project instructions. Someone uploaded three sample summaries as a style guide, which become reference files inside the skill, and a spreadsheet of this year's renewals, which goes to Project files. And if a developer once wired it to pull policy details from your management system, that's the connector you'll have to rebuild.

Sorting a custom GPT: procedure instructions, template and example files, and conversation starters go to a skill; standing-context instructions and client data files go to a Project; custom actions are rebuilt as a connector; the model setting, conversation history and sharing settings stay behind.

How to move one, step by step

1. Take inventory. List every GPT your team relies on, who built it, and whether it uses actions. The official record post has the two-column version of this, splitting the ones you built from the ones you only use, since only a GPT's creator or a workspace admin can migrate it.

2. Copy out what's inside. Open the GPT's configuration and paste the full instructions into a document. Find the original copies of its knowledge files. Write down its conversation starters, which OpenAI's migration FAQ doesn't list among what carries over, and note every action.

3. Sort it. Go through the instructions line by line with three highlighters: procedure, standing context, outside access. Most GPTs turn out to be one procedure wrapped in a lot of context. Some turn out to be two or three procedures that should become separate skills, and that's worth knowing too.

4. Build the skill. A skill is a folder with a file called SKILL.md at its center. That file starts with a name (lowercase letters, numbers and hyphens, up to 64 characters) and a description of up to 1,024 characters, under the open specification at agentskills.io. The description is the part people get wrong. It's what tells the AI when to use the skill, so Anthropic's guidance is to write it in the third person and say both what the skill does and when to use it. "Writes renewal comparison summaries for clients. Use when reviewing an expiring policy against a renewal offer" will get picked up. "Renewal helper" won't. Keep the main file under 500 lines and move long reference material into separate files. If your old instructions were one long prompt, the PRICE framework is a good shape for rewriting the procedure part.

In Claude, skills are available on every plan, including Free, and need code execution turned on. The built-in skill-creator will interview you about the workflow and build the folder, or you can upload a finished skill as a ZIP file. In ChatGPT, where your workspace has skills, you go to Skills, then Create, and choose to build it in a chat, in the editor, or by uploading one.

5. Rebuild the context and the access. Put the standing context into a Project's instructions and the client files into that Project. Custom actions don't survive OpenAI's migration, so anything your GPT did through an API needs rebuilding as a connector. For the big platforms, one may already exist in your AI tool's directory. For your management system, ask the vendor before assuming anything.

6. Test it against what you already know. OpenAI's advice after migrating is to compare familiar prompts and at least one harder case, and check that the right skill is picked and the right reference material is used. I'd make it concrete: take five requests you gave the old GPT all the time, plus one that used to trip it up, and run all six. If the skill doesn't fire on a request it should have caught, the description is the first thing to rewrite.

7. Share it last. Keep it private until it passes. In ChatGPT you can share a skill with people or groups in your workspace. In Claude, a skill you upload is private to your account until you share it, and on Team and Enterprise plans you can share skills to your organization's library.

A skill travels better than a GPT did

The skill format was developed at Anthropic and released as an open standard, and the list of tools that support it now includes Claude, OpenAI's Codex, GitHub Copilot, Gemini CLI and Cursor. A GPT only ever lived in ChatGPT. A skill that's mostly well-written instructions can go with you, which is a quiet benefit of doing this migration properly rather than just pressing the button.

Grade that one honestly, though. Each product adds its own extras, and the more a skill depends on scripts or one vendor's features, the more work it takes to move. Write it once, then test it everywhere you plan to use it. If you want to see where this goes once you have a few skills, How to Turn Claude into a Digital Crew picks up from here.

What this looked like for my own Showrunner

The Showrunner is the assistant that preps The AI Hat Podcast, and I wrote its role before the first interview, ahead of the show's launch on May 19, 2024. As a GPT, it asked me for the guest's name and LinkedIn profile, did the research, and offered topics, titles, questions, an intro and a bio. I'd also taught it how I like to introduce a guest. Prep that used to take me two hours or more came down to about 20 minutes, and the show has now passed 120 episodes.

I've since moved it from a custom GPT to a Claude skill. The routine at its center, the part that takes a guest and works through the prep, is exactly the kind of job a skill is built for.

Where to start

Pick the GPT your team opens most often and sort just that one this week. Three highlighters and an afternoon will tell you more about what that assistant really does than a year of using it did, and you'll come out with a procedure you can hand to any tool you pick next.

If you'd like to know where your organization stands before you rebuild a whole shelf of assistants, the AI Adoption Reality Check takes about five minutes. I read every one and write back within two days.

Frequently asked questions

Can I move a custom GPT straight into Claude? Anthropic's help pages don't describe an import from ChatGPT, so it's a manual rebuild. Copy the GPT's instructions out, sort them, and let Claude's skill-creator build the skill from the procedure part.

Do I need a paid plan? On ChatGPT, yes. OpenAI lists skills for Business, Enterprise, Healthcare and Edu workspaces. Claude offers skills on every plan, including Free, and Gemini offers them without a subscription on personal accounts.

Should the files I uploaded to my GPT go into the skill? Only the ones that teach the job, like templates, examples and style guides. Anything about a specific client belongs in a Project's files, because a skill is built to be reused and shared.

What if my GPT used custom actions? They don't transfer in OpenAI's migration. Rebuild each one as a connector, and check first whether one already exists for that system before you pay anyone to build it.