Devs, how do you use LLMs today while building software? There's so many models and different tooling, it's hard to keep track of everything. There's also a lot of hidden advertisements in LLM related articles for specific platforms. Therefore, I'd love to hear directly from devs how you use it today.
Discussion
i just use codebuff.com and ask it to do things for me, pretty simple. a lot of non bitcoin dev projects don't need much.
then i use claude for understanding existing codebases, by copy pasting the output of `$ tree src/` or individual files, or for fixing small bugs like missed curly braces etc
I solely use Replit Core with Claude 3.5 Sonnet as the main model.
I primarily use the assistant as a way to refactor code and help parse and fix error messages. The agent is great for quickly getting a UI and base foundation out for a new and simple app, but with context specific and more advanced requests, it can be a bit lackluster. The assistant picks up the slack where the agent lacks, however, since it has full context on the codebase and essentially performs the same functions as the agent (just on an existing project rather than from scratch).
Never heard about Replit Core before and tbh I don't really get what it is by reading on their website. What is it?
It's a cloud-based IDE that's basically an all-in-one app builder. You have the built-in AI (similar to Cursor) along with one-click deployments, native DB management, a bounty platform, etc.
Claude sonnet 3.7
Seems like a lot of people like to use Claude. Have you used any other models, e.g. OpenAI's latest models, DeepSeek, any Llama models? If so, how do they compare for software development?
I've used a lot of Chatgpt models as well, not quite as good as Claude, especially for frontend (imo)
no
there is ML stuff in my code editor but it is wrong about 75% of the time
only the autocompletion that reads actual grammatical context gets it right
but, then, i actually write algorithms, most of the time
i'm not some dumb scripter
you can write scripts in rust too, you know.
When I do a web search, sometimes I get an AI answer. In that case I read it but I am cautious to see if it is relevant, and I check it for accuracy by going to links below.
That's it. I don't use LLMs to program.
ive been doing trials of a few diff ones, ended up settling on windsurf for now.. (with mainly claude as the LLM choice).
its been blowing my mind, and i enjoy hearing about the cool stuff other nostr devs are doing with their agentic workflows. this stuff is changing the game so fast.
ive been using it to help me add features to various projects, when its on a roll, i feel like im shipping 10x faster.. then other times, it is not very useful and gets in the way. still learning how to drive this ferrari, but enjoying it very much.
Windsurf looks exactly like VSCode in the screenshots on their website. Is it basically that with built-in AI support? How does Claude compare to other models you've used so far?
it is built from VS code but their windsurf engine can edit code and chat with you, while loading up its context automatically.. in a side window.
claude is the LLM that ive gotten used to working well with code, i havent tried some of the latest chatgpt tho. the rest didnt seem to produce as good of solutions. i was on the fence about AI even being helpful till i started giving claude a run.
Heard from some ace devs - cursor is the best .. not sure if you can book up different models with it ..
I'm using Twinny in VSCodium with a self-hosted ollama, but I bounce over to DeepSeek or ChatGPT when I need more help. Here's an example of what can be done: https://youtu.be/EWvNQjAaOHw?feature=shared&t=4443
I don’t
programmers who use LLMs don't have programming skills
LLMs have programming skills like a 3 month seasoned javascripter
copypasta is all they got
I pay for chatgpt, use it everyday in my personal projects and at my work, use it for almost for everything from building frontend code html/css/ts to help explore the different options/stack/approaches to implement something, much like talking/brainstorm with another dev, when me and the got are on the same page on a project I just ask him for almost full feat code.