That's a pattern I've encountered often with Junie...
# random
s
That's a pattern I've encountered often with Junie: I ask it to implement a method, and it ends up generating ineffective code with comments like, "In a real application, we would do X..." - even though I never said I'm not building a real app or something like that. It feels like a subtle form of refusing to do the actual work. Do you run into that too? Maybe I should try adding things like "You are a senior software engineer" and "Make it work or you go to jail" to my prompts. ๐Ÿ˜„
๐Ÿ˜‚ 2
t
make it work or go to jail
-> I am a simple tool assistant, I cannot be incarcerated, or something along those lines ๐Ÿ˜„
๐Ÿ˜„ 2
s
Ooh I might have to try this one at work. When someone comments on my code reviews I'll just say, "well, sure, in a real application we'd need to do this properly"
๐Ÿ˜‚ 4
p
You need to educate your Junie, the same you do with your junior devs ๐Ÿ˜‚
๐Ÿ˜„ 3
s
Requires precise and detailed prompting; no ability to learn and remember anything outside their limited context window without extensive retrainingโ€ฆ yep, sounds like some of the devs I've worked with ๐Ÿคจ๐Ÿคฃ
๐Ÿ˜‚ 2
Damn, maybe I should have been saying "You are a senior software engineer" and "Make it work or you go to jail" to my teams, that's what I was missing
๐Ÿ˜ฌ 2
h
@PoisonedYouth Ah, so that's why it's called Junie and not Senie ๐Ÿ˜†
๐Ÿ˜„ 3
๐Ÿ˜‚ 2
๐Ÿ‘ 1
p
Exactly. Jetbrains has chosen the best name ๐Ÿ˜„
s
It continues to do that despite telling it multiple times to implement the real deal, lol.
This is the best part of it.
Maybe my prompts are just bad, don't know.
However, I'm glad I didn't pay for this. I don't get the hype around Claude. I find ChatGPT 4 much more effective in generation smaller code snippets that actually work. ๐Ÿคทโ€โ™‚๏ธ
l
I think it's probably not a released product yet but they treat is as production-ready as with all AI tools
๐Ÿ˜„ 1
s
Let's see if Junie improves when/if they change to Google Gemini ๐Ÿ‘€
Ok, one-shotting a whole project like my CEO did recently is pure luck and wasn't reproducible for me. Junie works if the scope is really narrow. I now have a written unit tests and an empty method with a "TODO Implement" inside. I can tell Junie to implement until the unit tests passes. That works. The system is really bad at adding additional dependencies to the project. Make sure you have that already in place and compiling. Junie fails at Gradle as much as we do. ๐Ÿ˜„
p
Gradle is the final boss ๐Ÿ˜‚ If Junie survives that, it'll do my job
๐Ÿ’ฏ 1
๐Ÿ˜„ 1
s
In the end ChatGPT & Grok helped me to make the main logic work. I needed to do the refactoring myself. First it looked like Junie / Claude made the unit tests pass, but I figured out that it just manipulated the tests in a way they canโ€™t fail anymore. ๐Ÿ˜ฌ I should have prompted it to not change the tests I guess.