It's cloudflare. It's always cloudflare

you only end up in a loop if you don’t define what you want clear enough. clarity has outcome, acceptance criteria, boundaries and non goals. once those are clear to you and you transform them into text, you have significantly better chances of avoiding the infinite loop.

tldr: you only get stuck in an infinite loop when you don’t know what you want(when to stop)

Design. And redesign.

and specs, and tests, and smoke tests, and more specs, and updated specs.

And don't forget processes!

And reading reports, endlessly.

Now please please make for with the first comment of every HN front page posts ever.

edit: added “please”

“The relationship of thought to word is not a thing but a process, a movement from thought to word and from word to thought.” Vygotsky

Assumption: now everyone can do more of the above. The final line is still selling. So everyone will get to the sales part, FASTER. Triage will still happen at this stage, regardless of AI. You won’t be able to avoid this triage, regardless of how fast you get there.

I can't find a name to dig more but the "everyone will get" part is something strange to me. If everybody has the same capability increase, then what changed really ? some would even say it will increase the paradox of choice.. more offer, still the same amount of time to decide, or maybe more AI based decision to match the amount.. so less human understanding.

“everyone”. it’s there, it’s accessible, it’s “cheap”. acceleration will depend on the operator capability. if the final product will make a diff in the real world, it will ALWAYS depend on the entrepreneur, not the tools used.

By that logic, same can be said about code. So, ig everyone “gets there” equally fast, then nobody has an advantage

The difference is the destination. Me crunching code with llms vs you doing the same thing will lead both of us to different places(let’s say equally fast for the sake of the conversation). What will set us apart in the end is who will actually sell it, and that’s llm-independent.

I have tried all the different models, different harnesses, different prompts. It’s not this. The people who say this would probably say the same thing about slot machines, you see, you have to bet 5 lines after you get a cherry no wonder you aren’t winning!

What i meant was: your experience might not be representative of the coding models capabilities. At least based on the poor examples in the post, you’re just scratching the surface of learning how to extract value from them, and, as i understand, you gave up and shared your experience. Given that you fully understand how an LLM works(and you do) it might also be that you’re using it outside training data. Not all code is equally represented(quality/quantity) in the training data, like there’s not the same amount of latin text as opposed to english text there(it’s all gravy to an LLM). You also suggested that manual might be faster and better quality - in MY experience, this is just you (using search through prompts), using it wrong.

Edit: added gravy pun

Do you pay your employer when you introduce bugs? I think you're lucky if you get usable output which you don't consider a mistake. Also, you might be mistaken if you think that you pay for a deterministic service.

edit: typo

Any accuracy measurements?

I majorly compared it to the native Explorer agents (for example in claude code). So far it has won against the explorer agents in 98 of 100 cases. I am already in the works to create a bigger benchmark, but did not have so much time for it. But you are welcome to test it out :)

I'm sticking with Sonoma as long as they still provide security updates, then i go +1, hoping to skip the whole liquid glass phase.