Gemini 2.0 Flash-Lite aces manual text, fails red-box overlays

The test is a concrete example of where multimodal AI helps and where it does not yet. The developer, who builds tools like ClickCapture and Robocopy Scheduler for Windows, wanted to avoid the manual grind of screenshotting, drawing arrows, writing step text and rebuilding parameter tables. But when the same model was asked to place red highlight boxes around UI elements, the boxes landed slightly off target or covered text, forcing a fallback to manual layout. The practical lesson is to split manual work into AI-doable and human-doable tasks, rather than expecting one-click full automation. That boundary may shift as models improve, but for now the developer recommends letting AI do the extracting and the human do the pixel-level polish.