Video Link Notes

1. One-command default:

bash {baseDir}/scripts/generate_note.sh --url "<video_url>"

This entry already does the app-like retries:

• local subtitles first
• remember and reuse the last successful cookie source
• auto-detect a fixed cookies.txt path
• audio download and transcription fallback

2. Local/open-source subtitle preflight:

bash {baseDir}/scripts/fetch_subtitle_via_ytdlp.sh --url "<video_url>"

3. Verify unified transcript resolver:

bash {baseDir}/scripts/resolve_transcript.sh --url "<video_url>"

4. Legacy desktop-app route is optional only:

curl -fsS http://127.0.0.1:8691/api/get_all_providers >/dev/null

Generate Notes

Quick hybrid entry:

bash {baseDir}/scripts/generate_note.sh \
  --url "<video_url>"

Optional flags:

• --backend hybrid|local-bilinote (default hybrid)
• --profile fast|balanced|quality (default fast)
• --cookies-file /path/to/cookies.txt
• --cookies-from-browser chrome

Default profile guide:

• fast: use local base + beam_size=1, best for OpenClaw day-to-day speed
• balanced: use local small + beam_size=2
• quality: use local small + beam_size=5

For OpenClaw-triggered runs, prefer:

bash {baseDir}/scripts/generate_note.sh --url "<video_url>" --profile fast

Do not call generate_note_via_local_backend.sh and do not probe the desktop app unless the user explicitly says to use the legacy BiliNote desktop route.

By default, the skill does not auto-read Chrome/Edge/Brave cookies on macOS, so it avoids unexpected Keychain popups. Only use --cookies-from-browser ... when you explicitly want that behavior.

For MiniMax note synthesis, the easiest setup is to put the key in:

~/.openclaw/minimax.key

The file can contain just the raw key on one line. openclaw.json remains a fallback source.

For a permanent app-like setup on Bilibili, prefer putting a Netscape-format cookies file at:

~/.openclaw/bilibili.cookies.txt

After that, keep using only:

bash {baseDir}/scripts/generate_note.sh --url "<video_url>"

The skill also remembers the last successful cookie source in:

~/.openclaw/bilinote.cookie-source.tsv

By default, remembered browser-cookie sources are not replayed automatically on macOS, so unexpected Keychain prompts stay avoided.

The skill now also caches successful transcript packs and final notes under:

~/.openclaw/cache/bilinote

This makes repeated runs of the same video much faster. To bypass cache for a one-off retry:

export BILINOTE_DISABLE_CACHE=1

Subtitle-First Flow

Primary path:

1. Try local subtitles first:

bash {baseDir}/scripts/fetch_subtitle_via_ytdlp.sh --url "<video_url>" --json

2. Or use the unified resolver:

bash {baseDir}/scripts/resolve_transcript.sh --url "<video_url>" --json

This keeps the app-like convenience layer while staying independent:

• subtitles first
• remember-and-reuse cookie source first
• automatic fixed-cookie retry for Bilibili
• local audio download only as last fallback
• audio transcription after download when subtitles are missing

How To Respond

When this skill is triggered:

1. Run resolve_transcript.sh --url "<video_url>" --json.
2. Read the transcript pack it returns.
3. Produce the final answer yourself using the transcript pack as source material, or call the bundled note synthesizer.

Return:

• a short summary
• structured markdown notes
• key points / key quotes
• action items if applicable

Do not rely on shell scripts alone to write the final note when transcript material is available; the agent should do the final synthesis.

Bundled synthesizer:

node --import tsx {baseDir}/scripts/synthesize_note.ts --input /tmp/transcript.json

It reads MiniMax config from ~/.openclaw/openclaw.json and falls back to heuristic notes if model calls are unavailable.

Output Contract

• Return:

1. A short summary.
2. Structured markdown note.
3. Action items (if applicable).

• If task fails, include backend error and the retry command.
• When subtitle fetch works but higher-level summarization is unavailable, still return a transcript-based note instead of hard failing.

Failure Triage

If the user explicitly wants the legacy desktop path and BiliNote returns 下载音频失败 or [Errno 32] Broken pipe, diagnose before blaming the provider/model:

bash {baseDir}/scripts/diagnose_bilinote.sh

What this usually means:

• The failure is happening in BiliNote's internal downloader stage (yt-dlp / ffmpeg / subprocess IO).
• If the same error shows up across both bilibili and youtube, it is probably not a single bad link.
• If 127.0.0.1:8691 is unreachable, the desktop app is not running.

Transcript-First Fallback

When local subtitles are unavailable, or when the user wants a download-first fallback:

1. Download audio directly, bypassing the BiliNote backend:

bash {baseDir}/scripts/download_audio_direct.sh --url "<video_url>"

2. Prefer independent local transcription backends, in this order:
   • whisper CLI
   • mlx-whisper on macOS / Apple Silicon
   • faster-whisper
   • reuse an existing local faster-whisper model directory when available, including the bundled BiliNote model path

Default quality/speed tradeoff:

• bilinote now defaults to profile fast, which maps to Whisper base for better speed in OpenClaw.
• If you want faster but rougher results, set BILINOTE_WHISPER_MODEL=base.
• If you want better quality and can accept slower runs, try BILINOTE_WHISPER_MODEL=medium.
• For a faster CPU path, the skill now defaults to faster-whisper with int8, explicit CPU thread count, beam_size=2, and VAD enabled.
• If you want even faster runs, set BILINOTE_WHISPER_BEAM_SIZE=1.
• If you want to preserve more decoding quality, set BILINOTE_WHISPER_BEAM_SIZE=5.
• If VAD cuts too aggressively on pause-heavy audio, set BILINOTE_WHISPER_VAD=0.
• To pin CPU usage manually, set BILINOTE_WHISPER_CPU_THREADS=<n>.

For Bilibili audio fallback, local transcription now defaults to zh unless you override:

export BILINOTE_AUDIO_LANGUAGE=zh

Recommended setup:

bash {baseDir}/scripts/setup_local_asr.sh mlx-whisper

bash {baseDir}/scripts/transcribe_audio_fallback.sh --audio /path/to/audio.m4a

3. Turn the transcript into a note with this structure:
   • 3-sentence summary
   • structured markdown note
   • key quotes / key points
   • action items

Use this fallback especially when:

• local subtitle extraction is unavailable for that link.
• The user only needs transcript + note quality, not screenshots or desktop-app features.
• You want a provider-independent backup route.

Safety

• Never print API keys.
• Prefer local/open-source subtitle extraction first.
• Only use local BiliNote backend when --backend local-bilinote is explicitly chosen.