[aDyslecticCrow]

This is a nice idea in theory. But you need excellent docs in the firstplace for it to work.

And if a human spent painstaking effort writing excellent docs, the least bit of respect i can give them is read it.

[andybak]

> But you need excellent docs in the first place for it to work.

Are you sure? I just tried it on projects of mine that have almost zero documentation it did a fairly good job.

[aDyslecticCrow]

Really? How large is your project?

There is a very clear point in codebase size where LLMs tend to falter without very clear written down overview descriptions of the system structure. I have a hard time seeing that this system would be immune to that.

i have encountered LLMs seeminly knowing more about a system than it should because there are many similar in its training set; but that just lead me to be extra sceptical when it pulls up functions that dont exist. (Ive fought LLMs about json libraries quite a bit)