Impossible to override wasteful and annoying built-in behavior.

Claude code embodies that branch of computer science that suggests that if you need to make a phone call, but don't know the number to call, your best first move is to invent a number, or compute an average telephone number, and call it. When that fails, try another. When that fails, back off to looking it up.

It's expensive in terms of time and tokens to have to endure this behavior.

The experience is made worse by the vacuous cheeriness ("Perfect!" -- when in fact nothing is perfect), and the constant evaluation of everything the user says, e.g. "You're absolutely right!" when in fact the system has not verified whatever the user said.

I have filled user and project CLAUDE.MD with copious "mandatory protocols" to block this behavior, but these are almost always ignored. It's built-in.

We all know that LLMs' major weakness is their need for proper grounding. The behavior I've noted here pulls in the opposite direction, almost flaunting that very weakness.