A two-layer pattern: a writer skill (this one) fetches a source, distills it, and stages a markdown file. A reader layer (you build) picks that file up and does something with it — reads it aloud, renders it as slides, pipes it into a dashboard.
This skill is layer 1. You bring layer 2.
What it does
Given a YouTube channel handle, this skill:
- Finds the channel's latest uploaded video
- Fetches its auto-generated subtitles
- Asks Claude to distill the transcript into two sections:
## Spoken brief— 2–3 sentences, voice-ready## Deck brief— topic + core claim + 3–4 proof points + closing beat
- Writes the result to
${MORNING_BRIEF_DIR:-$HOME/morning-briefs}/YYYY-MM-DD.md
Prerequisites
yt-dlp—brew install yt-dlp(macOS) orpipx install yt-dlpclaudeCLI — Claude Code, active subscription- Shell:
bashorzsh
Usage
/morning-brief-starter @SomeChannel
Pipeline
1. Find the latest video
CHANNEL="$ARGUMENTS"
CHANNEL_URL="https://www.youtube.com/${CHANNEL}/videos"
VIDEO_ID=$(yt-dlp --flat-playlist --playlist-items 1 --print id "$CHANNEL_URL")
VIDEO_URL="https://www.youtube.com/watch?v=${VIDEO_ID}"
2. Fetch the transcript
WORKDIR=$(mktemp -d)
yt-dlp --skip-download --write-auto-subs --sub-langs "en,en-GB,en-US" \
--convert-subs srt -o "${WORKDIR}/%(id)s.%(ext)s" "$VIDEO_URL"
TRANSCRIPT=$(sed 's/<[^>]*>//g' "${WORKDIR}/${VIDEO_ID}".*.srt)
3. Distill with Claude
PROMPT='You are producing a morning briefing from a YouTube transcript.
Return TWO markdown sections, nothing else:
## Spoken brief
(2–3 sentences, speakable, second-person. No creator names, no URLs.)
## Deck brief
Topic: <one line>
Core claim: <one line>
Proof points:
- <line>
- <line>
- <line>
Closing beat: <one line>
Keep it tight. No preamble. No apology.'
OUTPUT=$(printf "%s\n\nTRANSCRIPT:\n%s\n" "$PROMPT" "$TRANSCRIPT" | claude -p)
4. Write the brief
BRIEF_DIR="${MORNING_BRIEF_DIR:-$HOME/morning-briefs}"
mkdir -p "$BRIEF_DIR"
DATE="$(date +%Y-%m-%d)"
printf "%s\n" "$OUTPUT" > "${BRIEF_DIR}/${DATE}.md"
echo "Brief written to: ${BRIEF_DIR}/${DATE}.md"
Extension points
This starter does less than you may want. That's the point — ship the pattern, let you add the parts that make it yours.
- Add layer 2 (the reader). Anything that consumes the brief file: macOS
say, an ElevenLabs voice, a dashboard, a slide generator.awk '/## Spoken brief/,/## Deck brief/' "${BRIEF_DIR}/${DATE}.md" | say - Schedule it.
launchd(macOS),cron(Linux), or a GitHub Action triggered at 06:00. - Swap the source.
yt-dlpis the head here; replace it with RSS, a Substack, a podcast transcript. The downstream pipeline stays the same. - Add traceability. Write a sibling
.meta.jsonnext to each brief recording what the source was (channel, video ID, duration) so you can audit later. - Add a privacy filter. If your own version will surface things that shouldn't land in a shared file, add a grep-based abort step before writing. Keep the filter regex in a private file; don't commit it.
Hard rules (if you fork this)
- Keep the attribution header in the SKILL.md body. See
ATTRIBUTION.md. - Keep the output structure (
## Spoken brief+## Deck brief) — any reader layer will contract against those section headers. - Don't hardcode personal or client data into the skill. If you need a sanitizer, keep it separate.
Not included on purpose
- TTS / voice playback — you wire your own
- A slide or deck generator — the
## Deck briefoutput is the handoff point for one - Cron scheduling — the OS scheduler does that better than a skill
- A privacy / banned-words filter — every user's list is different
Troubleshooting
- Empty transcript — the video has no auto-captions; pick a different channel or skip the day
claude -pconsuming API credits unexpectedly — if you're on a subscription plan, stripANTHROPIC_API_KEYandANTHROPIC_AUTH_TOKENfrom the subprocess env so the CLI falls back to subscription billingyt-dlpcan't find the channel — double-check the handle includes the@and is spelled exactly as it appears on the channel URL