There was a post here that exactly describes the problem I’m having from May. It was Moderated, but closed after a note that said something like “Yes, this is an issue, will update if we learn something” and then the issue was marked as closed. I’m having exactly the same issue and I’m hoping for advice or on how to fix it. In short, 15 skills I had written that were being used on Classic and Codex just stopped being available to the threads. They are on the manage page, they show up if I type @[and the skill] but the threads can’t use them.
Troubleshooting already completed
I have already performed all of the following:
- Reloaded the ChatGPT desktop application.
- Completely restarted the desktop application.
- Logged out and back into ChatGPT.
- Tested through the ChatGPT web interface.
- Tested through ChatGPT Classic.
- Logged out of ChatGPT web, ChatGPT Classic, and Codex.
- Logged back in after signing out everywhere.
- Uninstalled and reinstalled
the skills.
- Re-uploaded
the skills.
- Published/shared the Skill through the workspace interface.
- Confirmed that
SKILL.md exists and is readable.
- Confirmed that
agents/openai.yaml exists.
- Confirmed that the YAML permits the
chatgpt product.
- Confirmed that implicit invocation is enabled.
- Confirmed that the Skills use distinct names.
- Repeatedly queried the runtime Skill catalog after each change.
- Deleted the only custom Skill the runtime could see.
- Confirmed that deleting it reduced the runtime custom-Skill count from one to zero.
- Confirmed that none of the other previously working Skills appeared afterward.
The issue occurs consistently across desktop and web sessions.
Signing out of every OpenAI surface did not refresh or repair the runtime Skill manifest.
Hoping there’s been a resolution since May?
Welcome to the Community, @Craig_Hoffman!
Thanks for your patience. Newly created or updated Skills can take some time before they are fully available in ChatGPT.
Could you please try again 20 minutes after the Skill was created, invoke it, and see whether the issue still persists?
We’ve also added your report to the existing investigation to help improve this behavior.
Could you please retry the affected Skill at least 20 minutes after it was created or updated? If the issue still occurs, please send to support@openai.com the following:
- The Skill name
- A HAR file and screen recording while reproducing the issue.
-Mark G.
Mark, thank you for the reply. Sadly, no amount of time made any difference. I’ve invested about 30 hours into trying to figure it out. I did find a work around. I took the skills and recreated them, then changed the name of the skill. Then, the chat threads can find the newly named skills. The content of the skill file did not matter, just the name of the skill. I recreated it by taking a trivial skill (equivalent of “Hello World”) and naming it one of the “lost” skills. It did not work. Took that same trivial skill, changed its name, and it did work.
It appears to me that some sort of name corruption occurred.
I invested the time to recreate all my skills, rename them and what not (and added a few more) and all the new ones are currently visible and working. I remain concerned these will disappear as well, but they have been up and running for five or so days.
I had already opened a ticket with support and have now sent them a .har file. I did a light inspection of the file - and it doesn’t show much - all these problems appear to me to be on the backend.
Thanks for digging into this so thoroughly, @Craig_Hoffman, and for sending the HAR file. The same Skill content failing under its old name but working under a new one is a useful clue. It points to stale backend state tied to the Skill name or identifier, rather than the package itself.
Renaming and recreating affected Skills appears to be the best workaround for now, though it clearly isn’t a proper fix. I’ve added this name-specific behavior to the existing investigation. Please send any further updates, especially if the renamed Skills disappear again, to the support email thread.
-Mark G.
Thanks - yup - happy to help. I completely agree with your suspicion - it appears to be a stale name registration on the back end. Also, great to converse with someone about it! Thank you for your time. Unfortunately I have two tickets open several weekswith support (this one, and another related my platform vs workspace backend connections) and the process has not yielded any helpful results yet. I’m hoping for some progress soon.
Yeah, I can see how frustrating that must be, especially with both tickets sitting open for weeks. Could you drop the case ID here for the ticket where you sent the HAR file? That’ll help us investigate further.
-Mark G.
Here’s the ID for the Skills Issue: 12082734. Sadly, I started that one after weeks of frustrating interaction with support on the first issue (still open and not resolved) for which I also sent a .har file for that issue too. Again, another back-end issued. My suspicion - provisioning conflict in migration process from personal Chat to Business Chat on the Developer’s Platform side.
I was a little terse when this one started…
-C
Thanks for providing the case ID, @Craig_Hoffman. I reviewed the full case, and you’ve clearly completed the necessary testing and troubleshooting on your end. We appreciate the time you’ve put into documenting this.
I also saw that you tried sharing the HAR file through a Google Drive link. Since support can’t access files hosted on third-party services, please attach the raw HAR file directly to your email reply. That will help the team rule out any remaining possibilities and confirm whether this is a backend or provisioning issue.
-Mark G.
Google Email won’t support the 50 mb of the file. I tried to zip, but it’s still over 25. Is there another way?
Hmmm, can you try to send the view-only (Google Drive) link to the email again?
-Mark G.
Resent. Can I send a Dbx link?
Yes, please send the full link rather than a hyperlink, since hyperlinks aren’t rendering on our end.
-Mark G.
Hi @Craig_Hoffman, I’ve escalated this to our specialized support team for further investigation. They’ll share an update here once they have more information.
-Mark G.
Great! I just tried again with a dropbox link. I also broke up the hyper link into pastable sections to avoid any hyperlink creation on the rendering side. LMK if that worked better.
Thank you, @Craig_Hoffman. I've handed this case over to our specialized support team. It would be great if you could attach a short screen recording showing you reinstalling and invoking the skill.