What problem does it solve? Putting a speaker's corrected words onto their own recording is error-prone: recognizer output contains mis-hearings, corrections change word counts so timings drift, and automated rendering can produce a result nobody reviewed. This Skill defines a workflow that joins the corrected transcript to the original per-word timings and renders subtitles only from a page the operator has seen and approved. ## Core Features & Use Cases - Correction-aware timing join: Maps each corrected word onto the time span of its mis-heard counterpart from the saved Deepgram response, handling cases where corrections merge or split words. - Operator-controlled editorial decisions: Requires explicit choices about which stretches appear on video, whether silences are cut, and how the burn is styled. - Reviewable HTML deliverable: Produces a self-contained HTML page with the video and CSS-animated words that the operator opens locally, adjusts by hand, and approves before rendering. - Deterministic frame rendering: Uses headless Chromium walking the page frame by frame with ffmpeg muxing, driven by frame number rather than wall-clock time. - Use Case: After transcribing a dictated video with corrected text in writing/<project>/dictations/<slug>.md, invoke /subtitles <dictation file> to produce a new video file with the corrected words burned onto the picture. ## Quick Start Run /subtitles on the corrected dictation file for your recording and answer the three questions about which stretches to include, whether to cut silences, and how the burn should look.
