fermionoid/video-study

Video Study

For a video, replay URL, or screen recording, start with the video workflow. Automatically use Computer Use when available to inspect the player, watch relevant intervals and capture frames. Existing captions or a local transcription tool can supply the words in parallel. If the user already supplied the transcript and visuals and only asks for layout, go straight to editing.

Choose the output the user asks for. Use a full transcript for 逐字稿, a sequence of visual steps for a walkthrough, and short timestamped explanations for illustrated notes. For an unspecified "截图配文字" request, make a chronological illustrated reading version with timestamps.

Use the HTML template for the layout. Start with the title and one line of event metadata, a short note describing sources, coverage and editing, and a clickable topic directory for longer material. Use the language of the transcript. For a short transcript, omit the directory when headings suffice.

1. Set the coverage. Use the supplied transcript or captions as the timeline. When several versions exist, align them by timestamp and surrounding speech before resolving discrepancies. Preserve the source file and work on a separate copy. If the requested range starts after the recording begins, state the actual range in the opening note. If only audio exists, obtain a transcript through an appropriate authorized media workflow first.
2. Edit for reading. Preserve the original sequence, claims, examples, questions, caveats and answers. Add punctuation and paragraph breaks. In a lightly edited dialogue, remove obvious stutters or duplicate ASR fragments when meaning is preserved. For strict verbatim, retain fillers and repetitions. Fix names, numbers and specialist terms only when evidence supports the correction. Mark unclear passages in place (e.g. [听不清] or [术语待核]); keep a timestamped review list if there are many. Describe the actual level of editing and verification.
3. Build the reading order. Name chapters after what is actually discussed. Group short ASR fragments into complete turns. Break long monologues at natural paragraph boundaries. Put the source timestamp and speaker at the beginning of each turn, and keep continuation paragraphs visually attached. Use original replay times after cuts or edits. If there are no timestamps, omit them. Use a neutral speaker label or 发言人待核 where identity is uncertain.
4. Place slides at the point of discussion. Insert a supplied slide after the relevant turn, with a brief caption and alt text. Crop screenshots to the slide itself without clipping slide content. Use a labeled appendix for supplied slides whose placement is unverified, if they belong in the deliverable. Use relative image paths and ship the assets with the HTML, or embed the images for a standalone HTML. Escape any source text inserted into HTML.
5. Deliver and verify. Use the user's requested format. When a shareable document is requested without a format, prefer a searchable PDF and keep the HTML as the editable source. Export with a browser's print engine using the template's A4 margins and print backgrounds. Compare chapter by chapter against the source, including opening and ending passages, turn order, and any uncertain or missing spans. Open the rendered PDF, verify selectable Chinese text and working directory links, and inspect representative pages including the opening, a chapter boundary, a slide, and the ending. Check for clipped slides, headings stranded at page bottoms, split captions, missing glyphs and unexpected blank pages. Re-render after fixes.

Use the template as a structural example. Replace all sample text, IDs, and the sample figure; include only sections and figures the source warrants. Use the supplied visual hierarchy as the default and adapt typefaces, colors and page size when the user requests it.