The Gemini AI in firestore really does a remarkable job but one of its biggest flaws is that it behaves like a bull in a china shop.
I have to instruct it every couple of minutes to stop just executiing code and follow a Discuss, Plan Implement on command logic.
The eagerness of the program to be good coding partner is commendable but it will very ofter change part of the code that it never needed to or were instructed to and then very often it would be neccesary to just roll back as it sometimes seem incabale of correcting errors in code that was otherwise working fine.
1 Like
When you engage with Gemini in Firebase Studio chat, you have the ability to change the engagement model (and Gemini model). However, this isn’t an option with using App Prototyping agent. Using the “Ask” mode will prevent code changes.
In App Prototyping agent, I have had success with explicitly explaining that I’m asking a question and not to make code changes until I said it was okay to do so.
1 Like
Yes the prototyping agent is like a retreiver that sees a squirrel 
Users definitely need better control over the Prototyping agent. It needs to follow a workflow of Idea–> discuss---->agree------> execute.
Currently there is no way of confirming if the Agent correctly understands its instruction or that the instruction given to the agent is correct.
I have tried giving the Agent instructions such as Do not code wait for my instructions to code, This works well for about 3 chats and then it tends to runway again and just start coding without providing me with the time to read what the implementation is
here is an example of the Agent not really understanding the instructions of do not code
