On-Screen Video Lookup: Text, UI, and Code from Frames

The transcript says "and then it breaks." The error that actually appeared on screen for two seconds never shows up in it. This skill locates that moment and samples frames densely enough to catch it. It reads the frame at full resolution and gives you the exact text, with a timestamp you can check.

What it does

• Finds the right moment first, using a transcript search or scene-change detection tuned for screen recordings, so it doesn't blindly grab a frame every few seconds
• Samples densely around the target, which catches toasts, tooltips and validation errors that last only 1-3 seconds
• Crops and upscales regions at native resolution, then copies error messages, code, URLs and labels exactly as shown, with [?] for characters it can't read
• Picks the most complete frame, such as the last frame before a cut for typed code or the settled state after a click
• Flags where narration and screen disagree ("click Deploy" vs. the button actually clicked)
• States coverage plainly: which ranges were sampled, at what density, and what was never looked at

What you get

A direct answer backed by timestamped evidence lines marked SEEN or INFERRED, plus a verbatim block of on-screen text. It also lists visual-only details and mismatches and gives a coverage note. If frames aren't accessible, you get the exact ffmpeg commands or the list of screenshot timestamps needed.

Best for

Engineers and QA leads triaging screen-recorded bug reports, developers copying code from video tutorials, and anyone who needs a step-by-step guide built from a demo recording. Also useful for researchers who need what a slide or dashboard actually showed, not what the speaker said.