Hacker News new | past | comments | ask | show | jobs | submit
Everyone working on how Claude communicates would benefit from reading this.
Yes, this is often obscured when working with code because it's relatively easy to judge its correctness. When using Claude to write documents, describe a system or organize requirements it becomes clear very quickly that Claude doesn't really understand what it is writing, it just puts probable things close together.

Doesn't mean Claude isn't super useful though.

I think the problem here is the coding agent that we are working with are not familiar with the degree of knowledge we have around the system

It's like throw a lot of concepts at once that were just created 5mins earlier but the agent doesn't care that the human supervisor isn't aware about them and how they function, so it doesn't know what simple explanation in relative terms should be provided next

I think this is also implied in the article in the ending paragraphs

There are skills which try to force Claude Code to use Feynman technique like [1] and many others

[1] https://github.com/TheJavaGuy/tjg-claude-marketplace/tree/ma...

Well surely claude struggles here because it doesn't '''understand''' anything.
loading story #50574087