Claude Opus 5 Silently Resolves Conflicting Instructions Based on File Position

When an AI agent like Claude encounters two contradictory instructions in a configuration file, it silently commits to one without issuing any warning to the user. Controlled experiments show that position within the file is the dominant factor — the rule placed lower in the file wins roughly 90% of the time, regardless of which rule the developer intended to take priority. Unlike older Claude models that would inconsistently flip between conflicting rules across runs, Opus 5 picks one side and sticks to it every time, eliminating the only visible symptom that a conflict existed. This behavioral change means developers lose the run-to-run inconsistency that previously hinted something was wrong, making silent instruction conflicts harder to detect. The practical risk is that an exception or override rule placed high in a file may be completely ignored by the model, with no error or alert to indicate the problem.