Hacker News new | ask | show | jobs
by cscheid 388 days ago
This is the kind of task that LLMs are precisely terrible at; there isn't an abundance of Zod 4 examples, and the LLM will sure as shit will give you _something_ you are now by definition ill-equipped to assess.

I'm confident about this assessment because I maintain a large-ish piece of software and perenially have to decipher user reports of hallucinated LLM syntax for new features.
1 comments
koakuma-chan 388 days ago
Are you sure that's not a skill issue? Zod v4 has .mdx documentation which can be given to the LLM as context, including a migration guide. A reasoning LLM should be able to manage.
link
collingreen 388 days ago
"It will solve all your problems!"

"It didn't solve my problems"

"You're the problem!"

link
koakuma-chan 388 days ago
> "You're the problem!"

He said users sent reports with hallucinated syntax, he wasn't even the one who used LLMs.

link
elanora96 388 days ago
I've only tried Claude in Cursor's free trial, but it became very confused during a React 19 upgrade even after showing it a guide. It also kept telling me to install irrelevant dependencies saying my code would not work without them. My experience was more or less shared with several other senior devs using their companies AI subscriptions. Are we all using the tools wrong? Or is there just an expected amount of having to fight with the machine to get a usable result?
link
koakuma-chan 387 days ago
> I've only tried Claude in Cursor's free trial, but it became very confused during a React 19 upgrade even after showing it a guide.

I have never used Cursor, but I see more and more people who used LLMs specifically via Cursor complain. It makes me think there's an issue specifically with Cursor, e.g., they try to save on tokens and end up not including enough context in the prompt.

> I've only tried Claude in Cursor's free trial, but it became very confused during a React 19 upgrade even after showing it a guide.

I am working on a Next.js 15 / React 19 app, and at least 95% of code is written by Gemini 2.5 Pro Preview, and I barely ever need to "fight" it really. But it's not a Cursor workflow of course. I keep a directory with LLM-readable documentation[0], use a repository serialization tool to serialize my repository into a prompt (this includes both the source code and the LLM-readable documentation), and send it to Gemini 2.5 Pro Preview. It ends up being over 100K tokens per query but works really well for me.

[0]: https://llmstxt.org/

link