We spent the last few days trying to break our own AI coach

We gave APEX conflicting follow-ups, new restrictions, old workouts, and the classic:

“Make it harder.”

The interesting part wasn’t making the workout harder.

It was making sure the AI didn’t forget what it had already been told.

A harder request shouldn’t erase a shoulder restriction.
A new limitation shouldn’t let an old workout sneak back in.
And sometimes the correct AI response is simply: don’t generate the workout.

That sounds obvious.

It wasn’t.

We found several edge cases where the AI technically gave a “good” answer - but the system logic underneath wasn’t good enough.

So we rebuilt those paths.

Still a lot to improve, but this is probably the part of building APEX I find most interesting:

making AI know when not to be helpful.

Curious how other builders test this kind of failure in their products.