yt-digest

name	yt-digest
description	Turn a YouTube URL into a vault-ready research digest — transcript, key learning moments with timestamp deep-links, and screenshots of visual moments (slides, demos, dashboards). Use when the user pastes a YouTube link they're researching, says "summarize this video", "/yt-digest <url>", "youtube digest", or asks to extract key points from a YouTube talk. Output goes to vault/research/youtube/<slug>.md with screenshots in vault/attachments/youtube/<slug>/. Source MP4 is auto-deleted after frame extraction.
argument-hint	<youtube-url> [--no-screenshots] [--tags tag1,tag2] [--slug custom-slug]

yt-digest

Process a YouTube video into a single markdown research note in your notes vault. Drop a link, get a digest that's faster to consume than the video.

Output contract

• Digest: vault/research/youtube/<slug>.md — frontmatter + TL;DR + numbered takeaways (each with timestamp deep-link + screenshot if visual) + collapsed transcript
• Screenshots: vault/attachments/youtube/<slug>/frame-MM-SS.png — embedded in digest via Obsidian's ![[...]] syntax
• Staging: /tmp/yt-digest/<slug>/ — video + raw VTT, deleted at end of run

Headless-safe: all inputs come from CLI args, no interactive prompts, deterministic output path so an orchestrator can read the result back.

Workflow

0. Parse args

Required: a YouTube URL (any form — youtu.be/x, youtube.com/watch?v=x, youtube.com/shorts/x).

Optional flags:

• --no-screenshots — skip video download + frame extraction (transcript-only, ~$0.01, ~5s)
• --tags a,b,c — add to frontmatter tags: list
• --slug name — override auto-generated slug

If no URL is provided, abort with a one-line error — do NOT prompt the user (this skill runs headless).

1. Fetch transcript + video

mkdir -p /tmp/yt-digest
META_JSON=$(.claude/skills/yt-digest/scripts/fetch.sh "$URL" "/tmp/yt-digest/staging-$$" ${NO_SCREENSHOTS:+--no-video})

Parse META_JSON (one-line JSON) for: id, title, channel, duration_seconds, upload_date (YYYYMMDD), slug, transcript_path, video_path.

Override slug if --slug was passed.

Set up paths:

WORK_DIR="/tmp/yt-digest/staging-$$"
ATTACH_DIR="vault/attachments/youtube/$SLUG"
DIGEST_PATH="vault/research/youtube/$SLUG.md"
mkdir -p "$ATTACH_DIR" "$(dirname "$DIGEST_PATH")"

If transcript_path is empty: YouTube has no captions for this video. Abort with: "No captions available for this video. Re-run with audio re-transcription (not yet implemented) or skip." Clean up $WORK_DIR before exiting.

2. Flatten the VTT

TRANSCRIPT_JSON=$(.claude/skills/yt-digest/scripts/parse_vtt.py "$TRANSCRIPT_PATH")

TRANSCRIPT_JSON is [{start_seconds, text}, ...] with rolling-caption duplicates removed.

3. Pick key learning moments

Read the full transcript. Identify 5-10 key moments that earn a bullet in the digest. Each moment is:

{
  "start_seconds": 754,
  "title": "Concise one-line insight (≤ 12 words)",
  "quote": "Verbatim line from the transcript that earned this bullet",
  "needs_screenshot": true,
  "rationale": "Why a screenshot would help — e.g. 'slide showing the 6-bucket framework'"
}

Selection criteria:

• Concrete frameworks, models, or step-by-step processes
• Specific numbers, data points, or benchmarks
• Counterintuitive claims with reasoning
• Moments where the speaker references something on-screen ("as you can see here…", "this chart shows…", "the formula is…")

Set needs_screenshot: true ONLY when the speaker is clearly referencing a visual (slide, demo, chart, code). Pure talking-head insights get needs_screenshot: false.

Anti-patterns — don't pick:

• Intros, outros, sponsor reads, channel housekeeping
• Filler restatements ("so what I'm saying is…")
• Repeated points (pick the strongest framing only)

4. Extract frames

If --no-screenshots was passed, skip this step.

Collect timestamps where needs_screenshot: true:

.claude/skills/yt-digest/scripts/extract_frames.sh "$VIDEO_PATH" "$ATTACH_DIR" 754 1230 1890

Frames land as frame-MM-SS.png in $ATTACH_DIR.

5. Assemble the digest

Write $DIGEST_PATH using this exact structure:

---
source: youtube
url: https://youtube.com/watch?v=<id>
channel: <channel>
title: <title>
duration: <hh:mm:ss>
published: <YYYY-MM-DD from upload_date>
digested: <today's date YYYY-MM-DD>
tags: [youtube, research, <user-supplied tags>]
---

# <title>

**Channel:** <channel> · **Length:** <hh:mm:ss> · **Published:** <YYYY-MM-DD> · [Watch ↗](<url>)

## TL;DR

- 3-5 punchy bullets capturing the video's thesis. Each bullet stands alone.

## Key Takeaways

### 1. <title from moment> · [<MM:SS>](<url>&t=<start_seconds>s)

<2-4 sentences expanding the insight. Reference the screenshot if present.>

![[frame-MM-SS.png]]   <!-- only if needs_screenshot was true -->

> <verbatim quote>

### 2. <next moment...>

...

## Open Questions

- 1-3 questions the video raises but doesn't answer, or things worth trying
- Skip this section if nothing concrete comes to mind — don't manufacture filler

## Full Transcript

<details>
<summary>Click to expand</summary>

[00:00] First line of transcript text. [00:08] Second line...

</details>

Notes on assembly:

• <url>&t=<start_seconds>s — YouTube timestamp deep-link. Use & if URL already has ?v=, else ?t=.
• Format duration as H:MM:SS if ≥ 1hr, else MM:SS.
• The screenshot path is relative — Obsidian's ![[frame-MM-SS.png]] resolves via the vault's attachment folder config. If that fails for any reader, use the explicit form ![](../attachments/youtube/<slug>/frame-MM-SS.png) as fallback.
• Full transcript block: timestamp each line as [MM:SS] so future search lands you at the right second.

6. Cleanup

Use Python rather than rm -rf — Claude Code's safety policy hard-blocks rm -rf even when explicitly allowlisted, which would break headless dispatch:

python3 -c "import shutil, sys; shutil.rmtree(sys.argv[1], ignore_errors=True)" "$WORK_DIR"

This deletes the source MP4 + raw VTT. The vault keeps only: digest markdown + screenshots (~2-3 MB total per video).

7. Print result

Final line of output must be:

DIGEST: <absolute path to digest>

Any orchestrator can parse this to surface the file back to the originating thread.

Cost

Component	Cost per 30-min video
yt-dlp transcript + video	$0
Model digest pass (~7K in / 800 out)	~$0.01–0.04
ffmpeg frame extraction	$0
Vault storage (digest + screenshots)	~3 MB

Source MP4 (~100 MB) is deleted in step 6.

Headless / remote dispatch

The skill writes to a deterministic path and prints DIGEST: <path> as its final line, so it composes with any orchestrator that can shell out — cron, CI, or a chat bot. The pattern worth building: a Slack/Telegram bot that runs the skill on a pasted URL and posts the path (or the rendered TL;DR) back into the thread. "Drop a link in chat, get a digest in your vault" becomes a one-message workflow.

Model note for unattended runs: in a headless claude -p run, a small/cheap model will sometimes describe what the skill would do rather than invoking the scripts (turns burned, zero tool calls, nothing written). Route the unattended dispatch to a more capable model. The interactive run is fine on the cheap model.

Failure modes

• No captions: abort cleanly per step 1 — do not silently produce an empty digest.
• yt-dlp 403 / age-restricted: print the actual yt-dlp error and abort. Don't paper over with cookies-based workarounds without explicit user request.
• ffmpeg fails on a timestamp: log the failure but continue with the other frames. The digest still ships; the missing screenshot is just omitted (don't leave broken ![[...]] syntax).
• Vault path collision: if vault/research/youtube/<slug>.md already exists, append a -2 suffix to the slug. Don't overwrite.