We Built a ChatGPT Plugin. People Are Still the Hard Part.

AI sucks at the actual people part.

It can write an event plan. It can't make someone answer an invitation, decide they want to help, or show up on Saturday. A very confident paragraph doesn't fill an empty shift.

We just submitted the Proximatic plugin for ChatGPT and Codex. It takes the event you describe in chat and helps turn it into a roster you can review and use.

I did most of this from my phone. The code, tests, packaging and even the YouTube demo were AI work. I directed it, reviewed the results and kept correcting things.

Getting Codex to build a plugin for itself was going on two days. Slightly ironic. Here's where the time went.

1. Start with distribution

Building software is getting easier. Getting anyone to find it is still work.

If someone is already explaining their event to ChatGPT, asking them to copy the answer into another app and rebuild everything is a stupid handoff. The conversation already has the date, jobs and shift ideas. Let them keep working there.

That's the distribution bet. OpenAI's plugin directory gives us another way to be found after approval and publication. It doesn't guarantee traffic or a featured spot. We still have to make something worth using.

2. Decide what AI should actually do

Supply the event notes, people, teams, available times and supported rules. The agent prepares changes. Proximatic's scheduler checks eligibility, availability, capacity and conflicts. You review the roster.

If nobody has said they're available, the model doesn't get to decide that Tuesday probably works. Two empty places stay two empty places. A tidy lie isn't a staffed shift.

The agent also can't accept an invitation, sign someone up or check them in on their behalf. Participants do that themselves.

3. Get past the first narrow version

The first package handled private event drafts. We expanded it to teams and imports, supplied availability, event setup, roster reviews, publication, participant messages and attendance records.

That meant updating the connection permissions, instructions, tests, listing and demo together. We went through several package versions getting them lined up.

The final server has 36 tools. They let you revise the event, inspect gaps and prepare the next action in the same conversation.

4. Make the real connection work

The private connection worked. The public submission needed another pass on OAuth and client metadata support. We had to test it through ChatGPT too. A local test passing doesn't make another system connect.

We revised the consent screen to name the organization and explain access. Participant email gets its own permission. New users can sign up, verify their email and return to the original connection request.

The plugin is Proximatic Pro only. That's the organization's subscription to our product. Existing AI task limits and email allowances apply; purchases stay in the web app. See pricing.

I want distribution that can support the product. For this integration, that means a paid organizer connection. Manual events, signups, rosters and attendance remain free. Participants don't need Pro.

5. Make “review” mean something

“Looks good” in a conversation shouldn't be a blank check to change a roster and email everyone.

You approve a saved proposal for the named organization. Publication and messages get separate reviews. If the people or availability change, you need a fresh proposal.

An approved availability change applied once. Retrying it returned the same receipt. A flaky connection shouldn't duplicate your work.

We corrected a tool annotation too. Replacing a prepared candidate changes saved private work, even if nothing gets published. The description and flags have to say what the tool actually does.

Three tools still have generic further-review notices. The portal says those don't block submission. Review is still pending.

6. Spend more time on packaging than you'd expect

The public listing showed the right Proximatic icon. The installed private copy showed a globe. We tried alternate asset paths. Still a globe.

We recorded the private display problem and moved on. The public icon was correct.

We downloaded the dashboard's saved package and compared it with ours. All eight files matched. After several uploads, it's worth checking which version you actually submitted.

7. Make the demo with AI too

QuickTime didn't save the new walkthrough. We switched to Chrome screencast capture of the actual consent screen and ChatGPT interactions.

AI made the edit, narration and timed captions. The roughly 110-second demo shows an event revision, coverage, an approved availability update and retry, and separate publication and message reviews.

The upload had a file-picker detour too. We checked playback through to the end and published English captions. Yes, even the demo became a small software project.

Watch Proximatic in ChatGPT: plan, review and organize events.

8. Test the boundaries, then submit

We checked five positive workflows and three negative cases through a matching private review copy with the public client's authentication. The unpublished submission couldn't be installed from the directory. The scheduling test reached the billing handoff; that reviewer run didn't execute a paid schedule.

We supplied reviewer access, the demo, test instructions and package. I authorized the legal attestations, and the submission went through. It's In review. Publication comes after approval.

Connections currently expire after ten minutes and can require reconnecting. The private icon still needs work. This was mostly built from my phone, but it wasn't one prompt and a victory lap.

The ChatGPT and Codex guide covers access and approvals. The AI event setup guide explains the event brief and web flow.