Show HN: CSS generator for a high-def glass effect (glass3d.dev)

It feels like almost every week a new AI dev assistant is being launched. Like GitHub Copilot and Tabnine did few years ago I feel like some of these tools are becoming a part of my daily coding life. That said, it’s becoming really confusing and hard to stick to one or few tools, whether it’s a VS Code extension (my preferred setup) or an IDE like Cursor or Windsurf.

Right now, I’m getting decent results from combining Windsurf and Copilot in VS Code. It helps with quick code rewrites and generation, and speeds up my commit cycle.

But I’m curious, what are others using? Have you found a setup that meaningfully boosts your productivity, especially with existing/legacy code, or you still prefer not using any code assistant, if so then why?

Comments (12)

jasonthorsness · 14h ago

Claude Code in the terminal has found the winning solution IMO. I have a multi-tab terminal with a session open in each tab and I keep cycling between them giving Claude tasks. I also keep VSCode open for diffing and making my own edits (with GitHub CoPilot enabled) and I also use GoLand for debugging.

If you haven't tried Claude Code yet, it takes like 30 seconds to install and get started and shows its value very quickly.

a_tartaruga · 11h ago

This is my setup too. It's nice to keep the Claude tasks on a remote machine with no important state so there's no chance of privacy leaks or in case you accidentally allow them to rm -rf ~

justinram11 · 11h ago

Although I think Cursor has the best UX (love how they queue up diffs that you then go through and can review/approve/deny easily while the agent still works), I've since switched to Claude Code + JetBrains IDEs as the sweet spot for powerful IDE + AI Agent.

Perhaps I just haven't found the right combination of plugins, but I can't for the life of me get VSCode code navigations (go to, find definitions, refactor method, etc) to work half as well as a JetBrains IDE (much less get a debugger working).

ch_de_sm · 15h ago

Ive been using a tool called AgentOne . Been testing it for few weeks now and it's working quite well for code generation, refactor, documentation, security scans and works well across the whole stack.

You have use your own LLM key (eg. Claude) though. The great thing about this is they have a live cost tracker so you only pay for what you use.

My favorite thing about using this is a feature called AutoMode. I can give a detailed prompt and it can build a complete app (create directories etc) without any supervision. By time I return, the app is done. The debugging effort is pretty low too

KaranSohi · 15h ago

link please?

ch_de_sm · 14h ago

https://www.iterate.ai/agentone

msgodel · 14h ago

I use qwen2.5-coder via neovim plugin, Qwen3 directly via llama.cpp, aider with qwen3 for big things that need context, and just copy and paste to/from claude using a special "heredoc" shellscript I wrote which lets me just put most of the project directly into the chat and copy changes back out in one operation.

They don't do anything I couldn't but they keep my momentum from dying when I get annoyed/distracted/frustrated which is a game changer.

KaranSohi · 14h ago

I've used qwen but not in the way you mentioned. How big is the qwen model you run, 8B? would also love to know the context window you use, as sometimes the models go out of context...

Thanks for replying this was really insightful :)

msgodel · 13h ago

Yeah I think I use the 8B one with aider. It has some crazy long context, 32k tokens IIRC.

arjuunan · 15h ago

Personally for me Windsurf has been my saving grace whenever I run into issues with a lot of my React projects. I have also had great results with Cursor but was looking for something that can not go overboard but still be able to handle a lot of the complex commands and tasks that I put out within the prompt. There are a handful of tools I still need to try but with AI there's always something new each week.

KaranSohi · 15h ago

Windsurf seems good to me too, Cursor sometimes generates the wrong code and when repromted to fix it does not always do a great job. But again, it really depends on the what type of project you are using it for :)

soapdog · 14h ago

I am not. They can all burn.