Guide Stage 5 · Scale
Migrate your team's GPTs to skills
Custom GPTs stop running on 11 December. Here's how to move a team's library to skills without losing the good ones, or keeping the dead ones.
63days until GPTs stop running
Every team that took AI seriously in the last two years built Custom GPTs: a proposal checker, a meeting-notes formatter, a tone-of-voice helper, and a dozen nobody remembers making. OpenAI is retiring them. The good news is that the replacement, skills, is the same format across ChatGPT, Gemini, Claude and Copilot. Move a GPT once and it works everywhere your people do.
What's changing, and when
Per OpenAI's announcements, as checked on 8 Oct 2026:
-
No new GPTs
Creating new Custom GPTs ends. Personal plans already can't create them.
-
GPTs stop running
Existing Custom GPTs stop working.
-
Enterprise extension ends
The last date for approved Enterprise workspaces.
OpenAI recommends moving GPTs to Plugins. In its migration, the GPT's instructions become a skill within the new plugin, knowledge files become reference files, and custom actions don't transfer. ChatGPT Skills (on Business, Enterprise, Edu and Healthcare plans) use the open Agent Skills standard; they can be shared with colleagues or published to a workspace library, and admins can install them for groups.
The same file works on the other platforms your people use:
ChatGPT
Custom GPTs retire; Skills use the open standard. GPTs stop on 11 Dec 2026.
Gemini
Skills replace Gems, on SKILL.md. Gems stop for business accounts on 1 Mar 2027.
Microsoft 365 Copilot
Custom SKILL.md skills in declarative agents, in preview for Frontier tenants: up to 8 per agent.
Claude
Agent Skills are native. Team and Enterprise admins can provision them org-wide.
Use the date as the reason, not the deadline
Don't leave this for the week before. A migration is the one moment everyone accepts a tidy-up of the AI tools they use, so it's the best chance you'll get to retire the duplicates and fix the ones people complain about.
What a SKILL.md is
A skill is a folder. Inside it, one file called SKILL.md holds the instructions: a short header the assistant reads to decide whether the skill is relevant, then the steps, rules and output it follows when it is. Reference files (templates, style guides, price lists) sit next to it in references/.
---
name: reviewing-proposals
description: Reviews a draft client proposal against our checklist and flags gaps. Use when someone shares a proposal draft and asks for a check or a second pair of eyes.
---
# Proposal reviewer
Checks a draft against the house checklist…
## Inputs
## Workflow
## Rules
## Output
## Guardrails
## Examples- name: lowercase letters, digits and hyphens, 64 characters at most, matching the folder name.
- description: up to 1,024 characters, saying what it does and when to use it. This is what decides whether the assistant loads it, so write it in the words people use to ask.
- The body: the instructions. Keep it under 500 lines; move long material into
references/. - references/: the files it reads, one level deep, each mentioned in the body.
What migrates, and what doesn't
Most of a GPT comes across. The parts that don't are usually the parts that made it useful, so check these before you call one done.
| In a GPT | In a skill | Status |
|---|---|---|
| Instructions | SKILL.md bodyOpenAI's own migration turns a GPT's instructions into a skill. Worth restructuring as you go: steps, rules, output. | Moves cleanly |
| Name and description | SKILL.md front matterThe name becomes a lowercase-hyphen slug (64 characters max). The description has to say what it does and when to use it, because that's how the assistant decides to load it. | Needs work |
| Knowledge files | references/ folderOpenAI moves knowledge files into reference files. Check each is current before it moves; most GPT libraries are full of stale PDFs. | Needs work |
| Conversation starters | Example requests in the bodySkills have no starter buttons. Keep the best ones as example requests so the description matches how people ask. | Needs work |
| Capabilities (web, images, code) | Whatever the platform switches onA skill can ask for a capability, but the workspace settings decide whether it's there. | Needs work |
| Custom actions (APIs) | Nothing, yetOpenAI says GPT custom actions do not transfer. Rebuild them as an app or connector, or as a manual step. | Doesn't transfer |
Convert one now
Paste a GPT's instructions below. You get a restructured SKILL.md (purpose, inputs, steps, rules, output, guardrails and a place for examples), a references/ README naming every file to move, and a list of what didn't migrate. It keeps the author's rules and adds nothing that wasn't there.
The five-step plan for a team's library
Converting one is ten minutes. Converting a team's library is a small project, and it goes faster as a project: one list, one owner per GPT, one date.
-
1
Inventory
List every GPT your teams use, not just the ones you built.
- Ask each team lead for the ones their people open weekly
- For each: owner, who uses it, how often, knowledge files, actions or connected apps
- Check the workspace's shared GPTs, not just your own
-
2
Triage: keep, merge or retire
Most libraries shrink by half here, and nobody misses what goes.
Keep
Used every week, has an owner, and people would notice if it vanished.
Merge
Three versions of the same thing. Keep the best instructions and one name.
Retire
Not opened in two months, no owner, or it only works because of an action that won't move.
-
3
Convert
Restructure, don't copy and paste. A GPT prompt written as one long paragraph becomes a skill with steps, rules and an output format.
- Run each keeper through the converter or rewrite it by hand
- Write the description in the words people use to ask
- Move current knowledge files into
references/; drop stale ones - Rebuild or replace anything on the "didn't migrate" list
-
4
Test
Same requests, old and new, side by side. You're checking it's no worse before you ask anyone to switch.
- Five real requests from the people who use it, including one awkward one
- One request it should refuse or hand back to a person
- Test on every assistant your people use; limits and settings differ
-
5
Announce
Tell people what's moved, where it lives and when the old one stops. Two lines each, in the channel they already use.
- Share or publish the skill to the same people who had the GPT
- Say what's different (no starter buttons; just ask)
- Name the retirements, so nobody goes looking for them
Update your AI policy
Skills are easier to share than GPTs were: a workspace library, admin installs for whole groups, the same file on four platforms. That's the point, and it's also why your AI usage policy needs a paragraph on them before the library grows back.
- Who can publish a skill to the whole workspace, and who reviews it first.
- What can go in references/: no customer data, nothing confidential the skill's users couldn't already open.
- Connections: anything that calls an internal system goes through IT, as actions did.
- An owner and a review date for every shared skill, so this tidy-up is the last big one.
The free AI usage policy generator drafts the whole one-page policy, including who to ask.
Questions
When do Custom GPTs stop working?
Per OpenAI's announcement (as of 8 October 2026), creating new Custom GPTs ends on 26 October 2026 and existing GPTs stop running on 11 December 2026. Approved Enterprise workspaces have until 11 February 2027.
What replaces Custom GPTs?
OpenAI recommends moving to Plugins, with the GPT's instructions becoming a skill inside the plugin. ChatGPT Skills (Business, Enterprise, Edu and Healthcare plans) use the open Agent Skills standard: a folder with a SKILL.md file and optional reference files.
Do my GPT's knowledge files come across?
In OpenAI's migration, knowledge files become reference files. If you convert by hand, put them in the skill's references/ folder and say in SKILL.md when to read each one. Check they're current first.
What happens to custom actions?
They don't transfer. A skill is instructions and files, not an integration. Rebuild an action as an app or connector your workspace supports, or turn it into a manual step, and involve IT before anything calls an internal API.
Will the same skill work in Copilot, Gemini and Claude?
The format is shared: Gemini skills, Claude skills and Microsoft 365 Copilot custom skills (in preview) all read SKILL.md. Each platform has its own limits and settings, so test the skill on every assistant your people use.
Should we migrate every GPT?
No. Most team libraries have a few GPTs people rely on, a few duplicates and a long tail nobody has opened in months. Inventory them, keep and merge the useful ones, retire the rest, and use the deadline as the reason.
Is it safe to paste a GPT's instructions into the converter?
The converter sends what you paste to an AI model through OpenRouter, routed only to providers that don't store prompts or train on them. GPT instructions rarely contain customer data, but take out anything confidential before you paste.