Ask HN: Do you prefer AI coding in an IDE or CLI? And why?

3 points by canxerian 135 days ago | 7 comments

I’ve tried both and ended up preferring a hybrid. IDEs are great when the AI understands project context, types, and refactors across files, but they can also nudge you into accepting changes too passively.

CLI feels more deliberate. You think first, ask precisely, and apply changes consciously, which helps avoid over-trusting the model. It’s slower, but the feedback loop feels cleaner and safer, especially for security-sensitive work.

Curious if others feel IDEs optimize for flow while CLI optimizes for intent.

canxerian 134 days ago | |

The IDE gives me a "human-in-the-loop" experience and conversely, I prefer this for secret sensitive work.

I agree with CLI feeling more deliberate. It could be the lack of "accept" gates (that you get with Cursor, Copilot etc).

billconan 134 days ago |

I prefer IDEs, they have better support for me to drag and drop my screenshots to the ai agent, my projects are ui heavy.

canxerian 134 days ago | |

I'm the same

delaminator 135 days ago |

I do a lot of AI coding

Receipts: https://github.com/lawless-m?tab=repositories

I started in Cursor - the tab completion is superb. As an assistant to the coder it is incredible.

But then I started to lean on Sonnet more and more. I expressed my ideas and they came alive.

As I got better at prompting I can code at the speed of thought. So I switched to Claude Code entirely. I’m a Max 200 customer now.

On Windows I still use VS Code because of the built in file browser and terminals don’t play nice with the Windows ecosystem. It helps to keep the windows organised. It’s also nice to have a few extensions like markdown renderer, graphviz viewer, png preview etc. I don’t use any of the IDE parts of the IDE.

On Linux I just use the terminal. I use i3-wm and have multi terminals each with a different instance of Claude.

I’m usually working on 2-3 codebases at a time and swap between them to give a new prompt.

I also run a Whisper server and have built a voice-keyboard program. It has some built in tool prompts “browse” launches a web browser, “desk 1” changes to virtual desktop 1, that kind of thing. It’s TheHand repo in my GitHub.

So I bounce from Cc to Cc and speak a new prompt, even to Windows as that’s via RDP and TheHand can type into it.

Then there’s Claude Code on the phone apps, so I can work on GitHub code with my phone wherever I am and not need a constant network - i have a terminal app on the phone too but it’s a bit scrolly - a phone emulating a serial console emulating a teletypewriter emulating a punch card system is so far beyond parody nearly everyone thinks it’s sensible!

I would switch to a Claude Code based editing system in a heartbeat, once someone build such a thing.

canxerian 134 days ago | |

This is a really interesting setup.

When you say "switched to Claude Code entirely" do you still make manual edits in a text editor is CC doing 100% of the coding?

Also, you mentioned getting good at prompting - is there anything you do that's not already well documented?

delaminator 134 days ago | | |

I stopped using Cursor entirely with 6 paid for months left on my subscription.

You’re right to questions the originality of the code though. I can’t answer to that. I specify at a high level.

In fact I don’t even do that, the spec is created from a chat discussion about what I’m trying to achieve. Opus in planning mode creates documents like this based on our conversation.

https://github.com/lawless-m/Cuthbert/blob/main/PROJECT_OVER...

But I’m not building CRUD all day long. I’m building things i couldn’t find.

If we were creating truly new algorithms we should be writing papers and books.

I think I am doing new work. I don’t think it can just be vibed into existence. I do have to tell it what approaches to take. Like “memoize those results” or try X and measure it.