The Two-Strike Rule

The two-strike rule governs what to do when an AI task fails: you get one rerun with a sharper prompt, and if it fails again, you stop rerunning and refactor — break the task into smaller pieces, move the judgment part back to a human, or fix the broken input upstream. Two failures on the same shape of prompt means the problem is the task, not the wording.

Why it matters: The gambler’s fallacy of AI work: the third attempt feels like it’ll land because the second almost did, and each rerun’s output is slightly different, which reads as progress. It isn’t — the right prompt is not one more iteration away. The rule’s other job is stopping the opposite reflex: bailing to do it by hand at the first failure, which quietly caps AI’s share of your work at whatever succeeds on the first try. Refactor before you bail; bail only after refactoring fails.