[parkersweb]

I’ve been using on a personal side project - but found that LLMs seem to be permanently confused over how to interact with pocketbase - to the point where I’ve even tried creating a Claude Skill to try and reduce the confusion.

Wondering if anyone else has had a similar experience?

[input_sh]

From the top of https://pocketbase.io/faq/

> If you don't have the time to at least skim through the documentation and you plan to solely rely on some AI tool, then please do NOT use PocketBase!

It's a niche little product that's alpha-level quality and changes frequently, I don't know why you would expect LLMs to be good at it.

[p2detar]

I tried to use LLMs to help me with server-side pb coding, but it was mostly a flop. LLMs don't have an up-to-date state of pb's API. 2 out of 3 times the LLM would give at least a hint how to go about something and the rest is me manually editing the code, reading the docs or looking at pb's source code. All in all, I consider it a nice "pair-programming" type of experience but one can't rely on LLMs to do the pb work for you.

[catketch]

This is what worked for me… clone the pb repo and stick reference to it in my agents.md. I put additional notes in my own addendum.md in the line cloned repo. ChatGPT-codex variants handle it nearly flawlessly and no issues with being out of date. I use the same pattern for all “niche” libraries

[Humphrey]

Not with Pocketbase - as I haven't found I've needed to look into the docs too much. But I have come across a whole bunch of areas LLM's seem to always answer incorrectly for. For example, ChatGPT has almost never corrected told me how to use the UI in Davinci Resolve.