Complaint: Junie is aggressively making changes! Disrespecting my prompts and preferences!
I've been designing with Claude - great experience but costly - and decided to try Junie for the past two days.
I have a context.md file that I give her when starting the chat. I say “read the context.md before we begin”.
What does she do? Magically reads some “next steps” document and starts implementing everything she found in there.
Then I added this to preferences:
**Step-by-Step Mode**: Always operate in a strict planning-first mode.
**Explicit Confirmation**: Do not modify any files or start implementation until the user has explicitly confirmed the plan.
She seemed to adhere to it when i gave it as a prompt.
Next day, I tell her to read context.md again with the above contents. She does, and returns back to me without making changes. What a relief. I give her a .cs file and ask:
suggest custom exception types to replace the InvalidOperationException in this file
a custom exception for each, or LunyServiceException and LunyLifecycleException? what makes sense?
Again, rather than answer my questions (!!!) she just goes ahead and starts implementing based on whatever she deemed appropriate!
“Brave” mode was off and she was said to “Auto”. The context file is in .claude/context.md but she won't just ignore its instructions because of the .claude folder?
How am I supposed to work with Junie when every time I risk her just ignoring me, doesn't respect the textual nor the UI preferences and just races ahead making changes to my project at will?
Please sign in to leave a comment.
It keeps getting worse. I tell Junie to refactor a class holding a
Dictionary<string, object>to create a new “Variable” class that wraps the object. My prompt:It then goes poking around the project sources due to some comment mentioning a “LuaValue” type and thus considers it as a template, creating a 150 line straightforward copy (!) of that LuaValue implementation rather than just a simple
Variable { public object Value {get;set;} }Oh and of course it made a unit test file too.
Junie is clearly incapable of thinking small!
It's true that Junie can be a little unpredictable from time to time, especially in complex environments. To get the best results, it is recommended to use Junie's guidelines.md where you can specify all preferences and this will ensure that Junie actually follows them.
You can read more about guidelines here: https://www.jetbrains.com/help/junie/customize-guidelines.html
And a repository of guidelines for different projects prepared by JetBrains: https://github.com/JetBrains/junie-guidelines/
When asking questions, Junie can sometimes misunderstand your intent and start implementation. In that case it's best to switch from "auto" to "ask" mode.
You might also find better results by switching to a different language model with better reasoning capabilities. You can do so in Settings -> Tools -> Junie -> Model.
I recently discovered that Junie doesn't care for things I told him. My whole team made this discovery. So this is no single user problem, take it seriously.
And don't downplay it like Damian Lech does. I can explicitly tell him he shouldn't do something anymore, and if he wants to do it, he explains to me why there's no alternative and doesn't do it. The first thing he does is exactly what I forbidd him to do.
Fix this! It's becoming increasingly impossible to work with Rider and Junie.
Thank you for reporting this issue.
What is the LLM model that you were using? May I ask you to provide screenshots of your conversation for analysis purposes? If you would not want to share them publicly, we can as well switch our conversation to personal ticket through form here: https://intellij-support.jetbrains.com/hc/en-us/requests/new?ticket_form_id=66731&product=ai_assistant .
Regards,
Junie Support Team