Hacker News Reader

Comment by ost-ing | Hacker News Reader

Is anyone actually getting a performance boon having these llms directly integrated into the editor? I find copy pasting blocks of code to chatgpt works really well to get help in isolation, especially if im doing something mathematical or algorithmic, or researching a new data struct or theory. Do I want it to run loose in my repository? Absolutely not

View on HN

Supermaven (which is owned by Cursor) in Intellij is fantastic - not perfect, but definitely speeds up boilerplate work. If you are e.g. doing a bunch of repetitive refactoring, after one or two iterations it will pick up the pattern and start suggesting it for you. Same goes for writing things like DAOs or resources, where you already have the pattern established for a different domain in your codebase.

I use Claude web to kickstart entirely new work or to attempt large refactors.

I guess it depends on the kind of work you’re doing. I use it for UI development and backend API stuff and it’s stupid how much work it does for me. I literally write a composer prompt, alt tab for 2 minutes, come back and it’s implemented a brand new endpoint with new models, GraphQL resolvers, working tests etc. It’s awesome.

As far as editor integrations go I've only used vscode copilot, and it's the method I've had least success with.

Aider is impressive but I can't use deepseek api at work so aider becomes too expensive using sonnet.

Which leaves me copying pasting into kagi code assistant as my most feasible and reliable llm coding tool at work. And its actually very good, but the copy pasting gets tedious.

I use Cursor and I get a lot of benefit from its autocomplete suggestions, but its composer is horrible so I never use it. The dream of telling AI to make changes on its own hasn’t arrived

It took me a while to make the switch from copy pasting in a chat to Cursor, but I do see an improvement in my productivity in a couple of ways. For context, this is doing front-end development and knowing what I'm doing. The benefits aren't there if I'm using a technology/library I'm not familiar with.

For the most part, I just type code the same way I use to but I get:

- an auto-complete on steroids

- the tab feature reminding me of impacted code I forgot to update after making a change elsewhere (big one as I easily get distracted).

I very rarely use the chat/composer. Usually I'm faster by going through files manually and making changes myself helped by the features mentioned above.

I find copy pasting blocks of code to be pretty worthless. It takes way too much effort to give it enough context for it to do anything sensible with my code, so it's usually better to just not try.

Cursor is the first time that I have felt that a chat-like UX could actually be useful for coding, because it gets the context for me. I still prefer autocomplete (and Cursor's autocomplete is very very good), but chat is actually occasionally useful for me in large projects. Without Cursor chat is only useful for one-off no-or-low-context scripts.

I’m not, honestly - it generally takes about the same amount of time, if not more, to carefully handhold the llm through properly implementing the feature.

At this point it’s really a question of how you’d rather spend your time - managing an AI? Or writing code?

Even though generally I prefer the latter, it’s fun to take a break and give the former a short occasionally. I’d say currently I let cursor do its thing for maybe 1 out of every 5 or 6 tickets, usually just for the sake of variety, or if I’m spinning my wheels and need to look at something to get started.

I don’t think it’s much better when I’m asking for a chunk of code, but the autocomplete is an improvement over not having it for me. Also I like being able to easily give it a relevant file from the code and having it follow my style/use the libraries I’m using etc. without having to type all that stuff out in the prompt. I think it’s worth having it in the editor, but it’s not as good as it needs to be for some of what they say it can do yet.

Reply on news.ycombinator.com