[embedding-shape]

> how do you trust the code claude wrote?

If that's something you're worried about, review the code before running it.

> don't you get anxiety "what if there's an error in tui code and it would mess up my git repo"?

I think you might want to not run untrusted programs in an environment like that, alternatively find a way of start being able to trust the program. Either approaches work, and works best depending on what you're trying to do.

[kaoD]

> If that's something you're worried about, review the code before running it.

It takes more, not less, time to thoroughly review code you didn't write.

[embedding-shape]

Depends. If I was the one coming up with the implementation anyways, it's basically just the "coding" part that was replaced with "fingers hitting keyboard" and "agents writing to disk", so reviewing the code certainly is faster, you just have to "check" it, not understand it from scratch.

If we're talking receiving random patches where first you have to understand the context, background and so on, then yeah I agree, it'll take longer time probably than what it took for someone to hammer with their fingers. But again, I'm not sure that's how professionals use LLMs right now, vibe-coding is a small hyped world mostly non-programmers seem to engage in.

[kaoD]

> you just have to "check" it, not understand it from scratch.

How can you "check" that which you don't "understand"?

> I'm not sure that's how professionals use LLMs right now

I'm a professional and I can tell you how I use LLMs: I write code with their assistance, they don't write code for me.

The few times I let Claude or Copilot loose, the results were heartbreaking and I spent more time reviewing (and then discarding) the code than what it took me to later write it from scratch.

[embedding-shape]

> How can you "check" that which you don't "understand"?

??? I do understand, since I literally just instructed it, how would I otherwise? I'm not letting the LLM do the design, it's all me still. So the "understand" already exists before the LLM even finished working.

> I'm a professional and I can tell you how I use LLMs: I write code with their assistance, they don't write code for me.

Hey, welcome to the club, me too :) I don't write code, I write English prose, yet nothing is vibe coded, and probably I'll end up being able to spend more time than you thinking about the design and architecture and how it fits in, because actual typing is no longer slowing me down. Yet again, every line is reviewed multiple times.

It's more about the person behind the tools, than the tools themselves I think ultimately. Except for Copilot probably, the times I've tried it I've just not been able to produce code that is even slightly up to my standards. It's a breeze with Codex though (5.2), and kind of hit and miss with Claude Code.

[coldtea]

>I don't write code, I write English prose, yet nothing is vibe coded

If the input is English prose, how is it not "vibe coded"?

[1718627440]

Maybe they are able to write unambiguous fully specified English prose? But yeah there should tell the world then, we others desperately don't know how to do that.

[coldtea]

>It takes more, not less, time to thoroughly review code you didn't write.

Nope, it takes way less. Else PR reviews would take as long as coding, which they obviously don't.

Writing 1000 lines, figuring out the nuances of the domain, fixing bugs, testing, takes way more time that reading and reviewing the resulting code.

Besides, you can even ask another agent to review it. Different brand of agent even.