Ask HN: How much better are AI IDEs vs. copy pasting into chat apps?

I tried to vibe code a web app with windsurf and found it appeared to do really well in the beginning. But I had to harsh the vibes in two types of scenarios it couldn’t understand or solve on its own: performance and css animation.

When I copy paste via ChatGPT I direct the model on what to do and retain responsibility and understanding of all code getting saved.

I do really like Cascade in windsurf because it does a decent job of parsing the entire codebase and finding the spot to make changes. But I think I’ll probably be most successful combining my approaches and using Cascade at a coding level not a “vibe” level

I have tried claude code on a few different size code bases and so far very impressed. The code it generates is usually in the same fashion and tone as the rest of the code base.

As a solo-founder of https://flowyBoard.com, I've used heavily Cursor and chatgpt. My app is very heavy on the front-end side and especially css animations. Here is my conclusion, Cursor is very good duplicate and extend your current structure in the project. So in the beginning you need to be careful about the solutions suggested by Cursor. When the problem is deep, you have to do tests and ask cursor multiple times to reform it. Chatgpt is good when the problem is deep and you know which part of the code is the problem.

To answer your question directly "it depends". There are a lot of different subtasks to coding. Defining requirements, sketching/discussing architecture, coming up with a technical plan, executing it, creating tests, etc. Different interfaces are better at different things. Cursor/Windsurf agentic flows can be great at well defined tasks that are directly actionable within the abstractions of the existing codebase. If you want to build new things, you need to be a little more methodical, and this can mean using chat interfaces to have a dialogue with AI.

I use RepoPrompt [0] + Cursor Tab Complete (since it's easily the best on the market), but I don't use the in-IDE cursor/windsurf agentic flows unless I have a small task I need doing where I can describe it perfectly. RepoPrompt is an underhyped context selection tool that includes

- A prompt formatter so you can easily dump it + prompts into chat windows

- A built-in diff format prompt template that gets auto-inserted so you can auto-apply the resulting changes.

- A unique "Chat" mode that's an API frontend for multi-turn conversations with dynamic code context, templated prompting, etc.

- An incredible "delegate edit" feature where the results of your conversation get sent to cheaper/smaller LLMs to actually enact. The cost savings and increase in coherency are unreal. It's not unusual for a single turn to generate something like 50 specific edits across a dozen files that are completely coherent.

Right now my main workflow is to discuss architecture with a context dump + prompt to o3, settle on a solution, build a PRD, give that to 2.5 Pro to create a more granular plan, make edits, then send it back to 2.5 Pro to delegate edits to Flash 2.0/DeepSeekV3/o4-mini. Not unusal to see it produce 20-50 edits across a dozen files with perfect coherency. Any fixups or minor things I do in cursor.

There are also the CLI agentic code tools. I find claude code to be like a more powerful version of the cursor agent flow. They're great for directed tasks that involve a lot of exploration, like understanding how you need to integrate with some submodule on the other side of the codebase.

[0] https://repoprompt.com/

I settled on gptel, which is an LLM package for Emacs. It's kind of a mix of both approaches:

1. You have chats right there in the editor, easy to copy/paste and manage context/history without switching to a browser. You can also quickly add files or portions of files to the context or remove them again.

2. You can choose which model you want to use for what, granted you have an API key.

3. You can quickly highlight some code and ask for a change to it, which along with managed context is powerful.

I tried auto complete again and again but it doesn't work for me. At first I think "yeah, that's what I wanted to write", and then I have to look closer to realise it's not, and that completely breaks my flow and concentration. I can always write some pseudo code and proactively convert it to real code, I like to be in the driver seat.

Context management is really central to my workflow, I manage it like a hawk. Models tend to deteriorate as context content increases, in my experience, so I really try to keep it narrow.

For that reason, and because our clients didn't sign up for their code to be sent to Anthropic et al, I _mostly_ use models like I would use StackOverflow, not to generate non-trivial code I'd actually use.

But having the chats in my editor is really invaluable for me. Powerful text wrangling features make a difference in both speed and motivation.

I use it pretty heavily with pretty much only the high-end models and pay about $15 per month.