Video Knowledge Ingest

Use this skill as the default cross-platform video → transcript → summary → local-knowledge workflow.

Quick start

1. 1. Run the bundled entrypoint:

1. 2. Read the JSON stdout for paths.
2. Send the summary back to the user from summary.md.
3. Keep the stored files in the local knowledge base; do not move them unless asked.

Default knowledge-base root:

• - INLINECODE2

What this skill invokes

Core tools in the normal path:

• - yt-dlp — resolve metadata, fetch subtitles, or download media
• INLINECODE4 / ffprobe — normalize audio before transcription
• bundled scripts/whisper-gpu.sh — local Whisper transcription using the workspace GPU venv
• INLINECODE7 — generate the final written summary
• local filesystem — persist transcript, summary, metadata, and index entries

Platform-specific notes:

• - YouTube: prefer subtitles when available; fall back to media download + Whisper
• Bilibili: often falls back to Whisper; the script auto-normalizes bilibili.com/... to www.bilibili.com/... and strips spm_ tracking params
• Xiaohongshu: usually no subtitles; expect media download + Whisper
• Local subtitle/text files: skip download and summarize directly
• Local media files: skip yt-dlp; go straight to Whisper

Workflow

1. Normalize the source

• - If the input is a URL, use the bundled normalizer.
• Keep YouTube timing parameters (t, start, list, index) but drop common tracking params.
• For Bilibili, force www.bilibili.com and remove spm_* query params.

2. Try subtitles first

• - Run yt-dlp in subtitle-only mode.
• Prefer zh.* and en.* subtitles.
• Treat subtitle download as best effort.
• If any usable .srt / .vtt file lands, continue with that file even if another subtitle variant returned a non-zero exit code.

3. Fall back to media + Whisper

If no usable subtitles land:

• - Download best audio/media with INLINECODE23
• Transcribe with bundled INLINECODE24
• If GPU transcription fails, the script falls back to CPU automatically

4. Summarize

• - Summarize the resulting transcript with INLINECODE25
• Expect codex to be installed and logged in, or configure the summarize backend another way before use

5. Persist results

For each ingested item, keep these files:

• - source.url or INLINECODE28
• INLINECODE29
• INLINECODE30 (when remote media/subtitles are fetched)
• INLINECODE31 (when Whisper was used)
• INLINECODE32
• INLINECODE33
• INLINECODE34
• global append-only index: INLINECODE35

Common commands

Remote URL:
CODEBLOCK0

Local files:
CODEBLOCK1

Custom output root:
CODEBLOCK2

When to read bundled references

Read references/toolchain.md when you need:

• - dependency details
• exact file layout
• how each tool is used in the pipeline

Read references/troubleshooting.md when you hit:

• - YouTube anti-bot / cookies issues
• Bilibili 403 on shared links
• subtitle 429 / partial subtitle failures
• Xiaohongshu subtitle absence
• summarize / codex auth failures
• Whisper venv, CUDA, ffmpeg, or yt-dlp problems

Operating rules

• - Prefer the bundled scripts/video-ingest.sh entrypoint over re-implementing the workflow.
• Do not skip the local knowledge-base write unless explicitly asked.
• When a run fails, inspect the generated directory before declaring total failure; partial artifacts often explain the real issue.
• If a platform provides subtitles, prefer them over Whisper.
• If subtitles are absent or unusable, fall back to media + Whisper automatically.