[Swizec]

> A theoretical “AI dev” will be able to gather these requirements through conversation no problem

Maybe! So far what I’ve seen are AI devs that are too afraid to say ”No you’re wrong this will never work” and can’t ask those awkward questions like ”Ok A sounds great, but you already promised to customer B that not-A will always be true. What do we do about that?”

It is very important to have a working mental model of the system to do requirements gathering effectively. AI struggles mightily in that area.

[jf22]

But all you have to do is tell the AI your mental model and it remembers.

And it will remember all the times some suit promised customer zyx four years ago this feature would be ready, and it remembers the insights the old PM had before they left, and it can read the four bug tickets related about a feature that describes conflicting behaviors...

[s__s]

Out of the box sure. An agent set up specifically for this task with your code and internal docs is a whole different story.

And LLMs aren’t afraid to say “you’re wrong”, they’ve just been configured to be agreeable by default.