The Point Where You Stop Correcting and Start Over

The correct answer is "/clear, then write one better prompt with what you learned."

Once the same issue has been corrected twice, the context is carrying several failed approaches that keep influencing the next guess — a third correction inside that same polluted session tends to compound the problem rather than escape it. A clean session with a single precise prompt that bakes in what the failed attempts revealed outperforms continuing to patch. Switching models or asking Claude to explain its own failures first are both still operating inside the same polluted context, which is exactly what the rule says to leave.