Hacker Newsnew | past | comments | ask | show | jobs | submit | Deeds67's commentslogin

Just create a skill for it -> I call mine `babysit`. It spins up a subagent that polls it every x minutes and auto-fixes it until it's green. I already continue with the next task while it does that in the background

I do this with our AI PR review checks. We have AI review every PR and commits to PRs... which can cause long running loops of commit<>fix.

So my agent just listens for green checks and no PR comments and loops until those conditions are met.


About 30% of our AI PR review checks are flawed at a fundamental level, based on poor comprehension of the whole. If I told the AI "fix everything and keep going until it's green" I would be terrified of the result.

I disbelieve this works in anything other than a toy codebase (or an incredibly fine-grained microservice).

The 70% is amazing! But a 30% failure rate requires intense supervision.


I'm building an open source Google Photos alternative. Have a look at my project and tell me if you think you could do all this at the same speed without an LLM: https://opennoodle.de

Direct github link: https://github.com/open-noodle/gallery


It doesn't completely blow away your argument, but you forked Immich and gave it to a LLM. Which is arguably slightly easier than starting from scratch.

Nothing wrong with forks though.


Sounds like you're nitpicking on whether or not LLMs make you faster by saying that working on existing code is easier than starting from scratch? I think you've missed the point.

No, they're saying that it was misleading. OP said "I'm building an open source Google Photos alternative" and surprisingly didn't say "based on Immich". This dramatically changes the evaluation of "open-noodle".

We're now in an era where LoC is easy and design is hard[1]. Starting with an existing project means using an existing design, where someone else has already made many/most of the difficult decisions.

10Xing code without caring about design/UX/DX is trivial. Literally anybody with a token budget can do it. But they probably won't ship a good project. Not with current frontier models.

[1]: design has always been hard. But now it's even more difficult because of code veloocity and because LLMs are happier to work with bad code than humans. It's never been easier to go deep into rabbit holes without noticing a single issue.


Pretty much every example in this thread is "I forked some existing project and made changes I like".

The main thing they dont realize is: 1. These are mostly superficial changes. 2. The only thing they 10xed is their ability to "start" on something. 3. They have not produced actual value. Their project/fork is just a version they think they prefer. But It is less maintainable, and less robust/useful for others due to its specificity.

My observations is that consistently these arguments are made by: inexperienced devs who simply dont understand what it takes to produce value in the real world.

LLMs CAN 10x you (in very specific areas like prototyping), IF you understand how to deliver this value, but that is the hard part. It has always been the hard part.


What are you debating here? I replied to a comment asking for examples of public work displaying that LLMs can 10x a human's output.

https://opennoodle.de/roadmap/

Look at what I built, these are not all simple designs.


Just that you're treating LoC and complexity as useful metrics, when these days those come for free. Or for $3 per Mtok, almost free.

To be honest, the official superpowers/brainstorming skill already does TDD so well, I don't see that much of a need for this. TDD is definitely the way to go with agentic development.


how?i saw superpowers/brainstorming but never saw tdd code produced


It’s supposed to do this, but I’ve found it doesn’t always do it


Just tell it to use TDD


There is another skill for tdd. You can activate it manually or tell the harness to

I'm the person working on that fork. Yes, it has now diverged 200k+ lines, but half of that is specs, research and documentation and includes a month worth of work.

The comment in question was a small feature of about 1.5k lines changed and it was solidly tested.


Eh, fair enough. 1.5k is reasonable. Have you tried just writing it yourself instead? It's likely it'll be less than 1k lines and you should have no problems writing an implementation yourself if you understand the structure of the LLM version.


Why would I write it myself? I use Claude code 12 hours a day and I'm really confident with want I'm able to build with it. I use it at work with incredible results. Spec driven development with harnessing is super powerful, I'll never be writing large features hy hand again.

Look at what I'm able to achieve in my free time after work https://opennoodle.de

Would've never had the time without my army of agents


Heh, fair enough. To me this comes off as "I'm unable to write it myself [possibly because I've outsourced my thinking too much]", to be honest, but I'm not going to argue; you're the one who presumably wants this code to end up in that repository.

I wouldn't really consider (what is likely) sub-1kloc a "large feature", but to each their own.


I don't want it to end up in that repo anymore, hence the fork. I've got a growing community of people who have been eagerly awaiting this feature and a ton more that I built.

I definitely could write this by hand - the stuff I built in the last 10 years before LLMs was more complex than this - but theres no way I'm spending all my free time to slowly craft something if I can just use AI and get the same results much faster


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: