Golden prompt set for ChatGPT Apps

ChatGPT Apps SDK
,
1

Does anyone have experience or advice on reliably triggering the model to display a widget? I’m trying establish a golden prompt set according to the directions on the “Optimize Metadata” page in the Apps SDK guide.

I can make my widget appear by typing its name into the text box, or by clicking the ‘+’ button and adding it. The correct tool is called, with the correct arguments, and the widget renders correctly.

But I can rarely make it show just by describing what I want to do, e.g. “I want to book a trip”. My resource and tool metadata are explicit that I support those intents (e.g. “use this tool to book a trip”), but nothing happens.

When I test my app with a tool like MCPJam’s “Test Cases”, it works. Just not with ChatGPT directly.

Just wonder if anyone else has this problem, or if I’m doing something wrong somewhere?

2

**I think that’s expected behavior, especially in developer mode.

I think “Developer mode” doesn’t take into account some of the in-chat discoverability that they discuss in the docs around the model being able to recommend or use an app even when it’s not added.

A good way to check is by not adding it and saying something like “Use the {name of your app} to do {action your app does}.”
**
It should search for it and tell you it doesn’t have access. Then you can add your app to the chat and run the same thing, it should work.

3

Thanks, I really appreciate your help! I’ll give that a try.

It seems like there’s 3 levels:

  1. Connect to the app in developer mode, then add it to the chat with the + button or by name, then type. This should trigger a relevant tool call.
  2. Connect to the app in developer mode, then just type and the model will call the tool if it’s relevant.
  3. Just type, and the model will suggest connecting to your app if the intent is relevant.

Level 1 works fine for me, and I don’t think level 3 applies in developer mode (or has even been rolled out yet).

Level 2 doesn’t work for me, but the submission process seems to imply that it should- there is a page asking for 5 positive prompts (that should trigger your app), and 3 negative ones (that shouldn’t). What are you supposed to put here, and how are you supposed to test it?