Practical prompting, backed by research

The paradox: maximum effort does not mean “the same thing, but more reliable.” It solves more tasks overall but loses some of the ones the low level solved on the first try. And you pay several times more. The “naive ladder” method lets you avoid overpaying for the maximum without losing success rate. An external check works as the on-switch for the expensive mode: you run at the lowest level, check the result from outside (tests, a checklist, your own eyes), and only on failure rerun at the maximum. In the study, the ladder matched a fixed maximum on success rate across all models and tasks. Median savings were 15–40%, counting the cost of both attempts.