A developer’s three-week record of a scheduled Claude Opus 5.5 assistant covered 14 active days, 23 runs including partial reruns, and 48 data commits. Six illustrated mistakes involved reading comments and login-protected pages, tracking contribution status and deleting an unfinished task. Mitigations included comparing actual timestamps, opening tracked items directly and labeling yesterday’s values when retrieval failed. User-controlled completion and reminders for long-unanswered work were simulations on historical records, not proven live improvements; the Duet comparison was not a hands-on performance test.
Back to Tech
Three weeks with a Claude assistant reveal reading and state-management failures
Separate completion and deletion authority, and label stale values when data cannot be retrieved.