2 Tries, Then I Stop Prompting

I used to loop: prompt, get a confident wrong answer, rephrase, get another confident wrong answer. Forty-five minutes gone on something a breakpoint would have caught in five.

The rule that works: if the second suggestion doesn't fix it, stop prompting. Open DevTools, add logs, reproduce the smallest case, then hand the agent the actual trace. AI is better at explaining real evidence than at guessing from a description.

The threshold is two. Not three. Not "one more try." Two. What's your cutoff?