Ask HN: What is your current LLM-assisted coding tool?

11 pointsHiPHInch9mo ago4 comments

Hello everyone, I’m testing and comparing various LLM-assisted coding tools, and I want to know which tool you are currently using in your daily development workflow. Here are some observations and questions I have:

1. Cursor and Windsurf

   - Both work nicely on local, but they use token-saving strategies:

     - With very long context, they may truncate important information, causing the suggested code to miss key details.

     - Even in normal scenarios, complex cases might exceed context or quota limits, interrupting suggestions.

2. “Roo Code” and API-based approaches

   - Directly calling paid APIs (e.g., OpenAI’s ChatGPT/GPT-4 API) works well but is expensive.

   - Some free or community APIs (open-source mirrors, community editions) can be unstable, rate-limited, or slow.

3. Augment Code - It’s said to be one of the most “intelligent” commercial products, but it’s also costly.

   - Many recommend its ability to rewrite, refactor, generate tests, etc., but for simple code completion, its cost-performance ratio may be lower than some smaller vendors or open-source plugins.

4. Refact.ai

   - Listed at the top on SWE Bench, it claims to support code refactoring, generating comments via LLMs, batch rewrites, and more.

   - However, it seems rarely discussed in developer circles. How well does it support?

Questions for the community:

- Which LLM-assisted coding tool are you currently using? (IDE plugin, standalone client, or API-based)

- What are the main reasons for choosing it? (e.g., cost, response speed, context length support, feature set, etc.)

- What pros and cons have you encountered during actual development? Specifically, how does it perform for debugging, refactoring, generating unit tests, automatic bug fixes, etc.?

- If you have switched tools before, why did you switch?

Thank you for sharing your experiences!

4 comments

jasonthorsness9mo ago

I keep CoPilot-based autocomplete turned on everywhere and occasionally use the inline chat with Claude Sonnet 4 but honestly for a lot of stuff I'm still stuck on copy/pasting from ChatGPT o4-mini-high. I've found I can get the best responses when I fully control the context and inputs independent from my workspace.

garbagecoder9mo ago

I've been trying blackbox.ai in VSCode the last little while. In my very limited experience it either gets it close to right on the first few tries or gets stuck in a kind of testing loop, making more an ornate tests that still miss one critical bug.

It's the first sort of "magic wand" coding AI I've used, where in the past I just would ask questions of ChatGPT or Gemini.

runjake9mo ago

I use Cursor and GitHub Copilot as “fancy auto complete”. On rare occasions I use their chat agent but I run into context limits and I generally don’t like LLMs going in and modifying code I’ve already written.

I also use Claude to ask general code questions, to explain something, give me ideas or code examples.

I am comfortably behind the state of the art.

bawolski9mo ago

Cline with Gemini flash, Gemini pro when flash isn’t enough to push something through

j / k navigate · click thread line to collapse