Where it consistently fails: anything involving the interaction between systems. If a bug spans a queue producer and its consumer, or the fix requires understanding how a frontend state change propagates through API calls to a cache invalidation - the model gives you a confident answer that addresses one layer and quietly ignores the rest. You end up debugging its fix instead of the original issue.
My stack: Claude Code (Opus) for investigation and bug triage in a ~60k LOC codebase, Cursor for greenfield work. Dropped autocomplete entirely after a month - it interrupted my thinking more than it helped.