lesson-clips

Esta página aún no está traducida; aquí tienes la versión en inglés.

Use when: a teacher, tutor or coach hands over a class recording (a 60-minute English lesson, a grammar class, a coaching session; screen share, camera, or both as two files) and wants it cut by teaching point rather than by length: “把这节课切成知识点”, “每个短语一条”, “cut this lesson into knowledge-point clips”, “today’s phrase clips with Chinese and English subtitles”, “顺便出一份学习笔记”. Not for a course master + episodes (longform-to-short, recipe longform-course) or plain N slices of a talk (longform-to-short).

Inputs → Outputs: lesson.mp4 (+ optional camera.mov of the same lesson) → per point out/<pNN>/<platform>-<orientation>.mp4 (title card “Today’s phrase · 今日短语”, the explanation with a persistent header, key-term cards at their first mention, captions with the target phrase highlighted, bilingual by default), a designed cover per platform shape, post copy, <pNN>.<lang>.srt/.vtt per language; out/recap/ (a mini card + the teacher’s own line per point); out/notes.md + out/notes.pdf (study notes); out/report.json (firstpass per export).

Defaults (format `lesson-points`)

State before the first render: “按「课堂知识点切片」默认：原速（语言课不加速），不加 hook，每条一个知识点 + 标题卡，中英双语字幕（目标短语高亮），每节课附学习笔记和回顾”. Language lessons are never sped up; no hook montage; no series labels. Student names / faces never go in: list lines that name a student before keeping them.

Steps

export PYTHONPATH="$VSTUDIO/lib:$PYTHONPATH"
python -m vstudio.project new lessons/week12 --recipe lesson-clips --input source=lesson.mp4 \
    [--input camera=cam.mov] [--param layout=pip] [--param subtitles=bilingual]
python -m vstudio.project run --dir lessons/week12          # stops at "审核知识点"

Or by hand (same modules the stages run):

Sync (only with a camera file): python -m vstudio.lesson sync --source lesson.mp4 --camera cam.mov --out work/sync.json. Audio cross-correlation of the two tracks; confidence < 0.3 = a guess, then the container timestamps are used (or pass --offset, camera = source + offset).
Plan: python -m vstudio.lesson plan --source lesson.mp4 --out work/points.json [--count 6] [--provider auto|none]. Transcribes once (shared cache), then finds the points: phrase, vocab, concept, correction (examples attach to the point they illustrate). The routed model (task lesson_plan) picks sentence ranges + kinds + glosses; the rules fill in and always work offline.
Review work/points.json with her (checkpoint points): kind, start / end, title, gloss (the meaning in the other language), terms. Set "edited": true after editing so a re-plan keeps it.
Render: python -m vstudio.lesson render work/points.json --source lesson.mp4 [--camera cam.mov --sync work/sync.json] --layout auto|screen|pip|band|camera --platforms xiaohongshu:full,douyin --subtitles bilingual|mono|translated [--to zh] [--mask sticker --mask-target camera] --recap --out out.
Notes: python -m vstudio.lesson notes work/points.json --out out/notes.md --pdf out/notes.pdf [--title "Small talk at work"].
Check: out/report.json lists firstpass per export (red items fail the stage gate); look at the title card, one term card and one bilingual caption frame before handing over.

Layouts (two files)

screen (the screen share only), pip (camera inset bottom right, default when a camera is given), band (vertical: screen band + camera band under it; horizontal: screen + camera column), camera. One master per canvas among the platforms (9:16 and 16:9 render separately), so nothing is cropped later.

Captions

vstudio.bilingual: the source line + the translation as a smaller second line in the theme’s secondary ink (over_ink2); translated shows only the translation. The translation goes through vstudio.llm task translate (its fallback chain); persona subtitles.glossary ({source: target}) is enforced and subtitles.term_fixes are applied first. If every provider fails the clip is rendered mono and the report says so. Per-language SRT / VTT tracks are written for YouTube / B站 uploads.

Privacy

--mask sticker|blur hides the tracked face (call-clips’ cat sticker) in the picture (all), the camera tile (camera) or the screen (screen); report.json warns when a face was found on under half of a part’s frames. Look at the frames before publishing. Pinyin / IPA lines are not drawn (gap).