alt Hacker News> If you want fixes/features, you'll prompt your LLM and it'll make it so.
I've been using the agents since the day they came out. This, does not work. It works if you want an app that looks like it works if that works without any testing. If you use the app in any way that isn't boiler plate it will break. Because you can't prompt ai with a simple feature no matter how small and have it cover all the important cases.
What these models do is the absolute bear minimum to satisfy the user even if it means cheating. And anyone who doesn't look thinks its miraculous. It really is not.
So, you absolutely can prompt it to add a feature, but you need to prompt it 10-20 times to get the actual feature built. If you don't know this yet rip.
And before we continue, also known that when you prompt for the next feature. There is a high chance the last feature you built will be altered or deleted.
I'm not a sophisticated LLM user, but nor do I have tricky usecases.
Opus 4.5 with a plan mode is enough for me: it asks clarifying questions to map out the design space, it covers fiddly parts, it generates good unit/integration tests.
I have seen it cheat by deleting existing tests, since it's single-mindedly focused on the problem I give it. I've not used an LLM-to-review-an-LLM, but when I've used an LLM to review my own work, it notices when there are unrelated changes in my commit/working copy.
Besides, all of your critique is how I've heard product managers speak of human developers! And a tool needn't be perfect for it to be useful.