Ask HN: Are new AI code editors adding superficial value? | Hacker News Reader

Ask HN: Are new AI code editors adding superficial value?

26 pointsdcreater15 comments

Specifically referring to Cursor and Melt. It seems to be the rage among the bay area AI community but they're basically just VS Code skins. Im pretty happy with VS Code both features-wise and UX, but I think now is the golden opportunity to fundamentally rethink the craft & work of human coding and build a new UI/App from scratch. Including NOT building on Electron..

View on HN

I've played with Cursor, it felt like all it has is a simple wrapper around Claude. It repeats what I've been doing manually before, namely copy pasting small code blocks into Claude and asking it to fix it. Cursor doesn't add much to that, plus it struggles sometimes when Claude says to insert something somewhere, Cursor fails to find out where that 'somewhere' is. The 'edit in text' function feels identical to Github Copilot, perhaps better in immeasurable 'quality' since Claude codes pretty well.

Somewhere out there is a far more 'native' way of coding with the AI together, where the LLM constantly sees the entire code base as you code; you don't code 'into the codebase', but you code 'into the LLM'. Hard to explain what I mean here...

You could vectorize the whole codebase and then keep having a conversation about it, update files in the vector, and so on.

The problem is if the LLM actually understands nuances and dependencies in the codebase structure.

For some languages and clear frameworks maybe so, but in many cases it could still misinterpret. Doable though imho to improve it.

Cursor for me has been a game changer, particularly in 2 parts of my typical workflow: post-research implementation and chore mass changes. Between this 2 I think I am about 30-40% more productive than with Claude alone.

For the first example, in a recent architecture task, I took ~6 hours researching options, and would have taken ~2-3 hours implementing my version. It took ~20 minutes with Cursor given I was able to give it my desired implementation very clearly and it applies the code immediately. I take most of those 20 mins reviewing the code, not even writing!

For the second, I mean stuff like renaming many variables that have similar but not identical names, or rewriting the way functions are defined/documented so it's the same for all, or adding type hints to all functions. Stuff that you need to typically do 1 by 1 and can't do Ctrl+Shift+F or multiline edit for. With Cursor, I make the first change and it simply knows I want to do something similar for the rest of my file/workspace. I just press tab 10 times and all the changes are done. From ~30-40 minutos to ~1 minute on a task that happens often.

Both of those have made much more productive (easily worth the price). As a distant third, having the ability to have Claude + Search has also been pretty useful.

I do think that there is a danger to the tool. It is great for senior programmers, but I think a quite harmful poison for juniors. While I go from large amounts of research to a bit of code reviewing, I think a junior will likely go from 0 research to direct implementation to 20 hours of bug squashing because they trusted it blindly.

Some thoughts on what I'd think would be requirements for a step function improvement in code editors for the new age:

- Human collaboration is a first class citizen if not the main design pillar. Instead of having add extensions for git, github etc. these need to be baked into the IDE. In-line blame, graph/UI for git history, diffs, team code editing etc.

- Basic code assistance utilities: Intellisense, Linting, Static Analysis, Formatting, Docstring Preview, Type Hinting, Dependency graphs, Code Metadata should again be key features that ships with the IDE as opposed to being enabled through extensions.

- Modern AI based code assistance: in-line code assistive comments based on review of what you wrote, tab autocomplete (this is covered well by the AI code editors/extensions), automatic documentation

- Implementation in C/Rust with effective/efficient usage of compute resources. Zed does do this

- Baking in project management/collab from JIRA/Linear: This would be the only way to keep human engineers collaboration not woefully slower than AI 'engineers' who would automatically bug fix, push, do PRs, review PRs etc.

Emacs or (Neo)VIM is your answer. You can have all of this with just a bit of elbow grease.

- Environment, Containers and package management baked in with nice UX. Python (the most popular language today) has a very fractured ecosystem with conda, venv, virtualenv, pipenv etc. Replace these with using uv+ruff (as they are FOSS) baked into the IDE.

Try zed.

I'm confused why people are giddy over Cursor given how polished Zed is

I am installing zed right now on linux because you said this but I see the Windows version has to be built from source.

Cursor I think has taken off because it couldn't be easier to try out.

Take out 90% of the windows users from bothering to try the software and that is really going to hurt people talking up the software online.

I've tried Zed, I like it, it's rigid and fast. But the following is a major drawback.

https://github.com/zed-industries/zed/discussions/6562#discu...

I'm unaware of any tool that can fully extract the entire codebase structure and provide it to an LLM. It would be ideal if the LLM could comprehend the complete codebase and suggest improvements based on the application's context. Currently, I'm using CoPilot in PyCharm, but it's not very helpful for writing actual code. I mostly use it for tasks like type conversions or explaining some logic.

Cursor is amazing, esp if I index docs in and use multi composer mode to generate files and code. Once I get a good working example in my repo I'm way faster with Cursor than without. It's not just a Claude wrapper, but you have to take advantage of all its features for the integrated experience.

There are some open source competitors coming out that I'm watching but Cursor is still top for me.

Maybe there is a lot more value to be created by fundamentally rethinking coding but anecdotally the absolute value provided by cursor is significant. Prototyping things in a new language or framework for the slow and lazy is now ~5x faster.

Reply on news.ycombinator.com