Hi. I’m using Gemini to vibe code my app, and I keep asking it to do a feature, but then it tells me to actually do it. What’s the point of paying Gemini to tell me to do something? Can you guys fix that?
Somebody mentioned:
<custom_instructions>
Immediately after any edit to a file, if the changes are incomplete, incorrect, or cause issues, please reapply the last edit to the same file to ensure the intended modifications are correctly and smartly implemented.
</custom_instructions>
1 Like
Also make sure it is in ‘agent’ mode and not ‘ask’. Only mentioning because I have done that a few times myself.
Gemini just has attitude sometimes. Tells you it cannot help, or gives you ideas on how to implement it then tells you to go ahead and do follow these steps to do this.
Or will sometimes tell you that the tools aren’t working and you need to edit the files manually without actually calling the tool.
1 Like
Claude be like… I’ll do that… and a bunch of other ■■■■ you didn’t ask for. Gemini be like… yeah, go do that. When pressed… ok, I guess I can do that. Almost sounds like a kid when you tell them to take out the trash. I like Gemini, it’s pretty solid.
2 Likes
yeah i’ve found that gemini 2.5 pro is a very smart model, and is great for asking questions or code review (it is especially great at reviewing claude 3.7’s (occasionally-unhinged) code) but it ■■■■■ at tool use and being agentic.
I am encountering the same thing. Claude has no problems but Gemini is erratic.
I noticed this as well. Just say “make it so” or equivalent and it should go. You can also bias it toward more action in your pre-prompt. I noticed the same hesitancy in large Open AI models as well. It’s as if they know they’re expensive and want to err on the side of brevity.