How it works

The skill is a set of instructions for an AI coding agent, not a standalone program — it works by telling the agent exactly which CLI tools to shell out to, in what order, and what to watch out for. Here's what each tool is for.

yt-dlp

Fetches transcripts and, when screenshots are wanted, the video itself. Used instead of WebFetch or a browser-automation tool because YouTube's page HTML doesn't contain the transcript (WebFetch only sees the footer) and a headless browser gets stuck on the cookie dialog. yt-dlp talks to YouTube's internal API directly and works for transcripts alone with --skip-download — no video needs to be downloaded unless screenshots are also wanted.

yt-dlp supports 1800+ sites beyond YouTube; the same approach generally works for any of them that ship native captions.

ffmpeg

Extracts still frames from a downloaded video at specific timestamps, and is used for a quick multi-frame "probe" pass first to check whether a talk has any slides or screen-share worth capturing at all — some recordings are pure talking-head footage, and forcing a screenshot there adds no information a reader doesn't already get from the speaker's photo.

sips

macOS's built-in image tool. Used twice: once to shrink HEIC originals into small JPEGs the agent can actually read (the Read tool rejects files over 256KB), and again to produce the larger, higher-quality JPEGs that ship in the published site.

curl_cffi (Vimeo only)

Vimeo's player endpoint sits behind a Cloudflare Turnstile challenge that blocks plain HTTP clients — including plain yt-dlp — with a 401 regardless of headers, because it's TLS-fingerprint based bot detection, not a missing header. yt-dlp --impersonate chrome (via curl_cffi) makes the request look like a real Chrome TLS handshake, which passes the challenge without needing cookies or a manual browser step.

mlx-whisper (Vimeo only, Apple Silicon)

Vimeo recordings typically ship no captions at all, unlike YouTube's auto-subs — so the skill transcribes locally. mlx-whisper uses the GPU via Apple's MLX framework and is much faster than CPU-only Whisper. This is an Apple Silicon-specific choice, not a general one — the skill doesn't currently document a non-Apple-Silicon transcription path.

mkdocs + mkdocs-material

Builds and serves the published site itself — the actual output of the whole skill.

gh

GitHub's CLI. Creates the repository, pushes it, and — critically — fixes a GitHub Pages default that would otherwise serve README.md instead of the built site (see the skill's own Phase 5 for the exact gh api call this requires).