Gemini 2.5 doing way more than asked for, always

Describe the Bug

I’m consistently having problems with Gemini to the point that I just can’t really use it at all in Cursor.

The preview version of Gemini 2.5 (exp) worked amazingly and was my go-to assistant. The full version is a dumpster fire.

Every prompt I’m finding myself going:

  • Don’t make a readme about the fix
  • Don’t add unnecessary features
  • Don’t write tests I didn’t ask for
  • Don’t attempt to make new type files or look through my codebase for other things to fix
  • Strictly respond to the prompt

No matter what I do, it’ll go absolutely off-track and do about 50 things, all things I didn’t ask for and then usually ignore 2/3 (if not all) of what I actually requested (usually fixes) and instead will go making new features.

Even in green-field development, it’s more rogue than usable. It still overlooks usually very obvious aspects – even when told about them, misses fine details, ignores prompts etc.

It also gets stuck in thought loops (sometimes for minutes at a time repeating the same thing). Other times it’ll rewrite a file 3-4 times or fail at tool calls repetitively.

At the moment, there’s actually few models which are reliable – o3 is too narrow and slow, Claude is okay, but previously was better. I had better success with older models.

Is the integration with models getting worse – it seemed more reliable previously? Is something happening on the backend?

Steps to Reproduce

Use Gemini 2.5.

Expected Behavior

Doesn’t go totally rogue.

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.0.0 (Universal)
VSCode Version: 1.96.2
Commit: 53b99ce608cba35127ae3a050c1738a959750860
Date: 2025-06-04T19:21:39.410Z (2 wks ago)
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.5.0

Does this stop you from using Cursor

No - Cursor works, but with this issue