When an AI chatbot gets a reply wrong, what should the repair screen show?

“That didn’t help” is an awful place to leave someone. They have already explained the thing once, and the usual chatbot repair is a fresh blank box or a cheerful rewrite. A recent UXBench preprint builds evaluation from real feedback signals and treats recovery after a bad answer as its own task. That seems closer to what users actually feel. The repair screen should keep the question and offer a few specific moves: “too long,” “wrong source,” “I meant this part,” or “start over without using that assumption.” If it guessed something, show that guess where it can be corrected. Otherwise a bad answer becomes two jobs: spot the mistake, then figure out how to describe it to the same system again. What would make a chatbot’s second try feel like it heard the first failure?