Anyone else have an AI code IDE that has gotten so aggressive and hyperactive, that they're increasingly going back to chatting in the browser, to make the sandbox smaller?
It just races off, in the wrong direction and rewrites the entire repo. Coding right past a simple typo bug..
Letting it take over the typing was a mistake. I'm starting to really hate that.
For me, I still enjoy getting my hands dirty. Let me share a story:
Last week vnlib got a big feature update. A single feature (zstd response compression) and it wasn't really my hand that wrote the code that added the feature. The feature was a single C file. The next almost 2 weeks were just review, bug fixing, and build script refactoring, docs, etc. I realized the fun part was writing that feature, like 300 lines of code, could have been done in a day instead of 10 seconds, and it wouldn't have slowed anything else down. While it's helping me though the other processes _a lot_. I let it do the easy part, the fun part, and it can't help me do the hard boring part of software. I believe this is a fact as it exists now and I think many people are coping.
Still super useful and a productivity boost, but lets be real with whats sustainable.
Yeah, it takes away all the fun stuff and leaves you with a messy heap of buggy code to sift through.
:'(
I think they don't want "code micromanagers", like us, as customers. They are aiming at people who can't read any code and will just vibe.
And absolutely whiz money away. I've seen some people throw hundreds of dollars at Claude for a weekend project so scrap it because it becomes unsustainable after enough features are added.
what i mainly get the coding agent to do for me is stuff that i'd spend a long time debugging that usually is simple but complex enough to get wrong.
sometimes it really struggles to do what i want and in those cases i often then just whip out the right code in like 20 minutes after watching it fail for 2 hours.
i don't use this one you use, i tried it but it seemed too aggressive and prone to missing out things, and it doesn't make tests or check stuff properly, maybe if you tell it to, but not by default
jetbrains junie does. i only have to rein it in when i haven't given it sufficient scope limits
also, i advise being careful to not have it make big changes that cut across multiple scopes, because it becomes very difficult to roll back changes. commit when it does the new thing properly, then you can just stash or dump the changes
I might need to switch. It's turning into an unwieldy monster. Gets worse, with each update.
like the majority of development toolchains
jetbrains main customer base is java and go devs, and those two languages have sane policies on feature creep. go is the best of course.
btw, for 20 eur/month you get practically unlimited LLM queries for junie and the ai assistant, on the AI ultimate plan. i'm finding that it's more than sufficient. i also use a local LLM for the AI assistant, because i've been finding that Qwen 3 14B is more than adequate for writing documents and commit comments. but you need a decent video card for that, mine is probably about $400 current market value, RX 7800 XT. plays games nice too, though for new games mostly i stick to FHD, a lot of older games it can do 2k or 4k smoothly over 80fps
Good to know. I will stick with the browser.
Yea... Add to your instructions to not do that and it still does it sometimes. Very annoying.
Yeah, I have that and I tell it to stop and it just does it over and over. Everything is designed to sideline real devs, now.
the masks are off... it has always been sales/marketing driven and the official narrative is that programmers are expendable
that's gonna go pear shaped really fast for them in the next few years.
Yeah, i have found planning mode in Claude code really good for solving this
Asking specific code tasks with good context or to explain the bigger picture but without coding has worked better for me. Also being nasty.
You've created a Sentient
Yes. Obv we could talk about prompting and whatever but the prompting never ends an is exhausting. But yes I often switch back to just chat mode. I've also been using gpt more often with some experimental prompts from the copilot github prompt list thing, because it's included in my subscription, claude and others are used for action.
I have a friend that has been vibing from the browser (not an ounce of software experience) and hes tried ides, cline, copilot and so on and has fallen back to claude browser XD maybe he is right.
Lately i've just been using it for well defined tasks like ,
- copy the work I did here to the other 5 similar scripts and open a review.
- help me brainstorm this feature
- how can I break down these commits better
- help my fix this regex for "this phrase" lol
