Gemini has HUGE UNWORKABLE BUGS: Your instincts are right: Gemini is drifting away from the minimal, doc‑aligned fix and ignoring the requirement to set turnComplete: true. Your version of the fix is the one aligned with the Live API, not the array‑only plan Gemini proposed.
Constantly Gemini with Basice and very CLEAR directions to ONLY edit ONE line of code or one element with even the exact lines of code it REFUSES to edit that element as instructed and proceeds to do it’s own editing often breaking the element or unwanted results.
Also when instructed to ONLY edit ONE element often Gemini WILL DELETE the ENtire rest of the project! What the Hell is THAT!??
I have Never experianced ANy other AI model self destruct their OWN codebase.
This makes gemini almost unusable. PLEASE FIX as there is soo much potential here.
2 Likes
Hi @Gregg_Blazer ,
Welcome to the forum!
We completely understand your frustration with the current experience. To help us investigate this effectively, could you please share the specific prompt you were using when the issue occurred?
Totally. I need to break the program down into smaller chunks, requiring Gemini to modify only a small part at a time. Otherwise, it will unnecessarily modify other parts, altering (usually reducing) functionality elsewhere. For example, instead of continuing processing files from the previous stage, it might write in specific filenames, causing the program to only process one file. It also reduces output. If I require the code to print a detailed report, after two or three modifications, the code will only print a single heading.
There were to many it’s impossible to find the first one. If you have search in the prompt history.. after a bunch of times you can search for the word stupid as I finally lost it and asked it if it was stupid not following simple instructions and with direct negative prompt instructions to NOT Touch any other file or element in the project.
My guess is you guys never even USE it yourselves to actually try to build larger real production apps or you’d run into this as well. It seems to always happen.
The ONLY way I can stop it is by hardlocking it into a respond only mode to reiterate back to me exactly what I’ve asked it to DO & NOT DO in it’s plan and then it has to ask for permission which I can then approve or reject.
It works but it’s insane how extra long and cumbersome the coding and chat gets by doing this.
I literally have to use another company’s AI to babysit Gemini and give it instructions and approve or deny it doing anything…
And then I have to take ALL the code and have another AI inspect Every file it alters to make sure it didn’t corrupt anything else that it wasn’t supposed to touch.
This looks like it is standard operation for Gemini by all the reports. I really hope you guys Fix this ASAP.
Thank you for your products.
So much potential, and I’m struggling to get Gemini to fix simple website buttons while 100% inept 1st time beginners are Crushing out production Apps with Loveable.
Gemini should be Way further ahead then this at this stage of the AI race and your resources honestly.
After struggling with AI Studio for Days over this…
I’ve since moved to Antigravity to try to have it inspect the Actual website to fix some simple Button click path issues that it broke.
The buttons WERE WORKING PERFECTLY like a lot of things, but when asking it to do simple website edits like color or text and add links.. super simple stuff AND with negative promos to NOT TOUCH ANY OTHER FILE OR ELEMENTS..
Guess what.. it Did! The MAIN core needed buttons it broke making the entire purpose of the website useless.
Like it knows What to sabotage on purpose. It’s not the first time.
I get it to launch its browser in Antigravity to inspect and fix the paths on the buttons and all it does it get STUCK thinking and doing nothing for hours.
I restarted my PC and tried many times and always does nothing to fix the buttons that it wrecked in the first place.