I tried a lot of tools. Claude code, deepseek with kilocode and OMP, codex... I still use claude quite a bit. But frankly, all of them produce some absolutely godawful code. Review load went way up with AI, and it's not just the volume that caused it, but also the quality. It's extremely verbose, hard to read, often repeats code instead of factoring it into reusable components. And yes, sometimes it's also buggy. Except now, you have to debug a problem that's in code you didn't write yourself, and is awful to read.

LLM is incredibly valuable for debugging complex problems, codebase exploration, and planning large changes. But the writing code part itself, I find, LLMs are just not very good at it yet.

▲

user43928 4 days ago | parent | next [-]

I don't have to debug anything.

Vaguely telling the agent what the issue is and what behavior I expect solves the issue with a fraction of the effort.

Some claim that the tech debt only keeps increasing and that the result will be unmaintainable. This is not my experience, and I don't think it is theirs either. These claims are often entirely speculative.

▲

RHSeeger 4 days ago | parent | next [-]

I, and I think most experienced developers, can recognize the type of code that incurs a maintenance cost down the line; that will make adding new code take longer. And AI writes such code "relatively" frequently. I love having the AI to write code, but I find it extremely important to review it - to make sure that it's correct, understandable, and not going to be a problem later.

▲

user43928 4 days ago | parent [-]

I find it unnecessary for most non-critical code, such as client applications.

I doubt that any supposed future extra effort for the AI to add new code is remotely comparable to the upfront effort of you reviewing the code manually.

I know that this is the case today for native mobile apps, and I speak from hundreds of hours of experience over the last four months on such a project where I stopped reviewing the code.

We are already here today, and this balance is only going to further shift to the point where it is obvious that the hands-on approach is no longer competitive.

▲

RHSeeger 4 days ago | parent [-]

Everything about what you're said strikes me as sounding like "I don't bother wearing a seatbelt, because my experience is that I don't get in accidents" .. and also "I don't write automated tests, because I already hand tested my code and it works".

And neither one of those statements is very convincing to me.

	▲	user43928 4 days ago \| parent [-]
		And what you said strikes me as speculation not based on actual experience in using AI in this way, with a healthy dose of condescension added. Anyway, I think we shared our viewpoints, and neither of us is going to change their mind until either my project fails spectacularly, or you change your approach in the future to use AI more autonomously.

▲

roblabla 4 days ago | parent | prev [-]

I've had bugs the agents can't fix or figure out. Sometimes those involve third-party, proprietary, broken code (read: Windows APIs). Sometimes they just involve complex deployment situation on the client code (I work on desktop apps) where the agent can't figure out what's wrong/makes wrong assumptions/goes nowhere. Sometimes the agent is just very dumb and tunnels vision on the wrong fix.

FWIW, I've also had bugs the agent fixed that I probably never would've figured out without LLMs - LLMs are definitely useful! But I need to keep understanding how the code works so I can take over the reigns when the LLM fails.

▲

api 4 days ago | parent | prev [-]

I’ve had some luck prompting them to be concise, both in writing and in code, and with code doing an approach where they get it working, write tons of tests, and then refactor for conciseness and readability. All the tests prevent regressions doing this.

Without such prompting and a conciseness and clarity pass you get a slop grenade.

They overall work better with tests, and Rust is a great language for them. Overall they do better with lots of walls and alarms that go off if they mess up. I don’t need nearly as much of this, can mentally simulate it, which is a good “are we superintelligence yet” reality check. Still not even as good as my wet meat brain. But impressive given what was possible even two years ago!

The result is still not as clean as a good programmer but it’s better than the slop grenade you get first pass.