Interesting approach using Gemini for checking whether the answer helped. I wonder how that would work if the response is hallucinated.

As an example: Stripe has sandboxes, I wanted to know how to create them via the CLI but found nothing in the docs. Asked the AI and it was like "Just do stripe sandboxes create <name>". The command does not exist and the whole functionality is not planned to be added. If another AI now reads this it could interpret it as answered successfully?

Just thinking - but I like the idea of the tool. Would love to use it but currently don't have any public docs to test with. Good luck with your project :)