Routing: Which Skill For Which Input

This plugin ships two complementary skills. Pick based on what the user handed you.

Local audio or video file → macwhisper-cli

An .m4a, .wav, .mp3, .mp4, etc. already on disk. Run mw transcribe — see ../skills/macwhisper-cli/SKILL.md.

URL (YouTube / Vimeo / any yt-dlp-supported site) → yt-dlp first

Even when the user asks to “transcribe” the video, audio transcription is usually the heavyweight option. For most “TLDR / summarize / teach me about / what does this say about X” requests, the yt-dlp skill’s subs-first workflow is enough:

1. yt-dlp --skip-download --write-description --write-info-json --write-auto-sub --sub-lang en --sub-format vtt pulls the creator’s description, chapter markers, and auto-generated captions in one shot.
2. Description + chapters alone often covers a TLDR without reading the captions.
3. If that isn’t enough, Read the .vtt caption file.

See ../skills/yt-dlp/SKILL.md.

Fall back to audio transcription (hand off to macwhisper-cli) only when:

• No captions exist for the video (older uploads, creator disabled auto-subs).
• Captions are visibly low-quality (heavy accents, music, specialized jargon showing up as garbled words).
• The user explicitly asks for a high-accuracy transcript (research, legal, accessibility).
• The source is an audio-only URL with no caption metadata.

In the fallback case, the yt-dlp skill provides the audio-download commands. Once the audio lands at /tmp/yt-<id>.m4a, macwhisper-cli takes over.

Direct audio/video URL (not a video-sharing site) → depends

If the URL serves a format MacWhisper accepts (.m4a, .wav), curl + mw transcribe is enough — no yt-dlp needed. The yt-dlp skill documents this direct-URL pattern too for completeness (and covers the ffmpeg conversion step when formats don’t match).

Composed flows

The canonical composed request is “teach me about this video — I want the TLDR, then I want to ask questions.” That’s one pipeline: yt-dlp pulls subs + description → agent writes the TLDR → follow-up turns Read the same files.

Both skills’ outputs go to /tmp/ following the shared output conventions, so the agent can hand off between them by file path rather than piping text through context.