Write the task brief for an autonomous long-running agent — an exact success predicate over the artifact, an explicit list of near-misses that don't count, enumerated failure modes, and a return condition — so hours of unsupervised work can't end in an answer-shaped artifact that solves nothing. Use before launching any multi-hour/overnight autonomous run, when delegating open-ended work to parallel workers, or after an autonomous run returned something plausible that didn't actually solve the problem.
