Automate the third successful run, not the first

ProblemOne good run is indistinguishable from luck.

TechniqueRun one to observe (where context was lost, where work was duplicated, what was misunderstood, which approval appeared unexpectedly, which check failed); run two on a different but representative mission to see whether it avoids the previous failure without a new instruction; run three with no intervention, recording completion rate, interventions, review loops, time to accepted result and cost per accepted result.

Why it worksIf the same mistake returns in run two, the correction never reached durable memory, which is exactly the thing you need to know before scheduling.

When to useBefore converting anything to a routine.