I have to admit.

LLMs can get stuff done.

That's fundamentally why so many developers feel compelled to use them.

But I absolutely hated the process of using them.

I'm supposed to say what I want (can’t even type it), create like a gigantic requirements file, pretend I've gone through them, think I've planned everything ahead, pretend I'm controlling the AI, sit back and let it do everything.

If it works. Great.

If it doesn't, I just tell it something went wrong and tell it to correct itself.

This process sucks and has not worked for me.

So I stopped using LLMs altogether.

Coding became much more fun, I have to say.

But I still want to use it when I do research.

For example, I wanted to create an API route that returns HTML in a framework that I've never used. (My toxic trait is to use a different tech stack in every new project and tell myself that I can pick it up in 15 minutes.)

I know exactly how to do it, but since I'm using a new framework, I don't know the exact implementation.

I could go through the documentation and look for it. And pray that the guy who maintains the documentation actually loves his life.

But I could just go to an LLM to tell it what I want, and get back the same result in less than 5 seconds (doesn’t always work tho…)

I like that. I just don't want an LLM to think for me.

What I have been doing is use ChatGPT/ Claude separately and copy the code to my editor.

But that just doesn't feel very efficient.

This has changed when I recently decided to try Zed + GPT 6 Luna.

People have been recommending Zed to me, but I didn't listen. I'm sorry.

If I have to describe why Zed feels so good, I couldn't.

Probably it's because it doesn't shout at me about how I should code or force me to use their agent.

With Zed, I can have everything in one window: an agent thread on the left, code in the middle, and occasionally the file explorer on the right.

I don't have to leave the editor, and I can get everything done.

Pair this with GPT 6 Luna model. The experience is really sweet.

Luna is fast. So far, it's accurate. It doesn't try to do more than what I tell it to do or think that it knows something better than me.

It's like a nice little coding assistant.

Plus, it's actually cheap. It's even cheaper than DeepSeek, which is just wild.

If you don't like your current setup and don't want to pay a lot of money just to code, maybe you will like this.

Fee from Anime Coders