Video Archive Workflow

Video download, metadata extraction, deduplication, subtitle handling, and archive organization for URLs from YouTube/Shorts, Xiaohongshu, Bilibili, and X/Twitter. Use when the user wants to inspect, classify, dedupe, or build a downloadable/archiveable video record from a URL, including sheet field mapping and folder organization.

Overview

Use this skill to turn a video URL into a normalized record for a sheet and archive system. Prefer deterministic metadata from yt-dlp, parse tags from page text when needed, and keep the workflow moving with safe defaults unless a hard stop rule applies.

Core workflow

1. Parse the input URL.
2. Identify platform, video_id, and video_type.
3. Check dedupe by video_id only.
4. Extract metadata from yt-dlp JSON first.
5. Clean title, description, and tags.
6. Resolve uploader and target folder.
7. Download or archive only when asked.
8. Write sheet fields in the required order.

Use the reference

See references/field-schema.md for the exact sheet field order, tagging rules, and platform-specific normalization rules.

URL parsing rules

• Identify platform from hostname.
• Extract video_id from the platform’s native identifier.
• Set video_type by platform convention when possible.
• If video_type parsing fails, default to normal and continue.
• Do not block the workflow for an unknown type.
• If the URL does not point to a video, stop and report that no video record can be filled.

Dedupe rules

• Use video_id only for dedupe.
• If a matching record exists, show a brief summary of the existing video and wait for user confirmation before proceeding.
• Do not require platform for dedupe.
• Do not auto-skip on an uncertain match.

Metadata rules

• Prefer yt-dlp -j or --dump-single-json as the source of truth.
• Save the cleaned metadata fields into the sheet.
• Leave unknown or empty fields blank.
• Do not write placeholder text such as (empty).
• Do not write raw JSON paths, download status, or command history into the sheet.

Tag rules

• Extract tags from title and description first.
• Strip tags from title and description before writing them.
• Do not translate text.
• Keep the first-seen order.
• Remove duplicates.
• Output tags as a space-separated string in the sheet.
• For Bilibili, also parse page tags from tag-link text when title/description tags are missing or incomplete.
• For X/Twitter, use the tweet text if available; if the item is not a video, stop.

Uploader and folder rules

• Use the uploader display name for the folder name.
• If uploader cannot be resolved, stop and ask the user before continuing.
• Do not fall back silently to uploader_id for folder naming.
• For Xiaohongshu, set uploader_id to the displayed Xiaohongshu number when available.
• For multi-uploader cases, keep uploader and uploader_id as comma-separated values in order.
• Use the first uploader as the folder root.
• Sanitize folder names to remove illegal filesystem characters.

Archive rules

• Keep the original filename style.
• If the source container is mp4, archive to mkv with lossless conversion.
• Keep the source file as 原名_source.mkv.
• Save the archived file with the original base name in the archive container.
• If a filename conflict exists, stop and ask the user what to do.
• Use the uploader folder directly under the main archive root unless the user asks for deeper nesting.

Subtitle rules

• Download and embed subtitles only when subtitles exist.
• Skip silently when subtitles do not exist.
• Record subtitle language availability in metadata.
• Do not repeat subtitle presence in the remarks field.

Field formatting rules

• Use Beijing time implicitly; do not label it.
• Format upload date as YYYY-MM-DD.
• Keep source container as the container field.
• Keep resolution in the format returned by yt-dlp when possible.
• Use units in field names where required: 秒, MB, fps, kbps.

Stop rules

Stop and ask the user when:

• uploader cannot be resolved
• the URL is not a video
• a filename conflict needs a decision
• the metadata is ambiguous enough to affect folder placement or dedupe

Output style

Return only the fields the user asked for when they request a sheet preview. Keep it concise and do not add extra narrative unless the workflow is blocked.