Developer uses same AI to write and review code, catching 3 self-made bugs

A developer building a Windows screenshot-translation app discovered that asking an AI to review its own freshly written code — by role-switching to a 'skeptical senior engineer' — uncovered three real bugs the same model had missed during writing. The bugs emerged in a fix for an OCR pipeline that was skipping white text on dark backgrounds, and included flawed shape logic, over-aggressive noise removal, and incorrect brightness measurement. The developer notes that writing and reviewing code engage different cognitive orientations: writing checks intent fulfillment, while reviewing questions whether the intent was correct to begin with. Crucially, not every finding from the AI review was acted on immediately; each was triaged as a real bug, a design choice, or a false positive before any change was made. The experiment suggests that role framing — not model capability — can meaningfully shift what an AI notices in its own output.