video-subtitle

<用户自制> 视频字幕：提取 YouTube 与 Bilibili 的字幕并依据实际读到的讲话文本作答。 Retrieve exact subtitle tracks and read scoped spoken evidence when an answer materially depends on a YouTube or Bilibili video and usable authoritative text is absent. A video URL alone does not trigger this capability.

Keep the answer in contact with the voice it claims to understand. Evidence belongs to one video, one returned track and the portion actually retrieved and read. A fetched artifact establishes acquisition; only read content grounds the answer. Use this package when the task materially depends on spoken content and reliable usable authoritative text is not already present. Metadata, comments, danmaku and another summary cannot silently replace it.

This Skill owns triggering, exact-track choice, video/part scope, chunk reading and evidence grounding. The enclosing learning/course or analysis owner decides the question, relevance and learning judgment; the host owns tool access, credential handling and actual personal-authentication boundaries. The local source names no downloadable companion Skill for these roles. See the dependency and owner table: Python and yt-dlp are execution dependencies, and the private connector formerly used for transport is not a standard or required public companion.

Concrete portable route

Use Python 3.10+ and yt-dlp on PATH; follow setup and examples. The bundled wrapper has no browser-cookie discovery, private connector, login adapter, or media-download step. Use it from this package directory:

python3 scripts/subtitles.py list 'VIDEO_URL' > listing.json
python3 scripts/subtitles.py fetch 'VIDEO_URL' --listing listing.json --language 'EXACT_CODE' --kind manual --output transcript-run
python3 scripts/subtitles.py read transcript-run/RETURNED_FILENAME --start 1 --count 120

Inspect listing JSON before fetching. Replace placeholders with the actual URL, one exact returned code, its returned kind (manual or automatic), and the returned file path. Provider codes are case-sensitive and may contain dots or hyphens. Never normalize, guess, translate, prefix-match or select all. Prefer a suitable manual track when present; identify automatic captions and their possible recognition errors. YouTube generated translations are excluded by the extractor setting; verify the track's provenance if the provider/extractor classification is unclear. Do not silently translate or substitute another track when the selected track is absent.

Fetch re-extracts metadata, checks URL/video/scope against the listing and requests only the exact selected track. Inspect current title, video ID, track, scope, file bytes/hash, retrieval date and companion metadata.json. Signed provider URLs exist only in a disposable temporary input, not saved metadata. Keep local output directories unique; the helper refuses artifact overwrites. Its metadata is provenance, never the transcript body. Read the returned actual file and compare the read hash with fetch metadata before attributing its content.

Scope and reading

Use explicit Bilibili /video/ URLs and ?p=N for a requested part; resolve short links visibly before invocation. The wrapper conservatively reports the requested page and unknown multipart page count/whole-video coverage. It does not establish that all parts were obtained. Verify the extracted title/identity against the requested part, and describe the answer as limited to that extracted part. Fetch other required parts separately using their explicit URLs. A scope/identity mismatch between listing and fetch stops adoption; list again and investigate. Do not transfer Bilibili multipart claims to YouTube.

Read bounded chunks, continuing from end_line + 1 until needed coverage is established. The reader returns 1-based ranges, total and remaining lines plus the file hash; for JSON/JSON3 it renders timed cues as separate logical lines. VTT/SRT remain raw timed text, retaining overlap and repetitions. Treat cue count and timestamps as evidence rather than assuming repetition means new speech. For whole-track summaries, structure or judgments read all chunks, including the ending, and inspect missing/empty cues or unexplained time gaps. Preserve a compact working outline; do not copy a full transcript into unrelated durable memory. For local timestamp questions read the needed range and identify that scope. If a read fails, state the last confirmed range and answer only from it. A complete file read is whole-track coverage, not proof of all spoken content: captions may omit speech, and Bilibili multipart scope remains unknown.

Failure and completion

No available track, extractor/network errors, access restrictions and incomplete readback are different boundaries. Report the failing operation and what usable evidence was actually obtained. Update yt-dlp or diagnose current official tool behavior for a concrete extractor failure; do not evade restrictions or claim subtitles do not exist merely because one route failed. This wrapper does not handle login, CAPTCHA, geo restrictions or private videos. Any other authorized source or explicit user-provided transcript must be separately identified and verified; never hide a switch to speech recognition, translation, danmaku or an approximate summary. Existing host permission, credential and personal-authentication boundaries remain with their owners.

Completion requires the same URL/video/scope across list and fetch, one exact listed code/kind, a nonempty downloaded artifact with provenance, matching read hash, and actual cue text read for the claimed range. The answer cites video, track kind/language and useful timestamps, and states material caption/part/read limits. Tool installation, a listing, a path or metadata alone is not completion.

Provenance

This original public adaptation retains the source method's exact-track, multipart-scope and chunk-read distinctions, replacing its private file-resource bridge and host audit ledger with the bundled standard yt-dlp wrapper. No private helper implementation, connector ID, cookie discovery or account data is distributed. yt-dlp is installed separately under its own licensing; video/subtitle rights remain with their owners.