| name | video-screenshots |
| description | Capture screenshots from a YouTube video at specific timestamps and compile them, with recap text, into a styled PDF. Use when the user wants frames/screenshots/stills from a video, a visual recap, a "screenshot at each timestamp", or a PDF of key moments from a YouTube URL. Trigger with phrases like "grab screenshots from this video", "make a PDF of key moments", or "capture frames at these timestamps". |
Video Screenshots → Recap PDF
Download a YouTube video and extract frames at exact timestamps, then compile
them with recap text into a styled PDF. Uses the plugin's shared access layer,
so it survives the po_token/DRM (403), n-challenge, and datacenter-IP (429)
walls — frame capture needs all three cleared, so read
resources/unblocking-youtube.md first
(cookies are almost always required).
How to run
Simple — just timestamps (captions default to the timestamp label):
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/screenshot_pdf.py <url> \
--timestamps 0,6:00,7:30,11:30,32:30,44:00 -o recap.pdf
Full control — a claims file with sections and captions:
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/screenshot_pdf.py <url> \
--claims claims.json -o recap.pdf --subtitle "Title — Channel · youtu.be/ID"
claims.json is a list of {timestamp, section, text} (timestamp = SS,
MM:SS, or HH:MM:SS):
[
{"timestamp": "00:00", "section": "Intro", "text": "Cold open."},
{"timestamp": "6:00", "section": "Topic", "text": "Main subject begins."},
{"timestamp": "1:01:00", "section": "Outro", "text": "Wrap-up."}
]
Recommended workflow for Claude
- Fetch the transcript first (youtube-transcript skill) to pick meaningful
moments — topic shifts, on-screen artifacts, deictic cues ("look here").
- Build a
claims.json with a short caption + section per moment; always
include the first frame (00:00).
- Run
screenshot_pdf.py. It downloads once, extracts every frame locally,
and renders the PDF.
Read a couple of the JPEGs in the work dir to verify they're real content,
then deliver the PDF.
Requirements
yt-dlp, and ffmpeg (or pip install imageio-ffmpeg for a static binary).deno + the po_token provider + cookies.txt — see the unblocking playbook.reportlab: pip install reportlab.
Notes
- Downloads a compact H.264 copy so the static ffmpeg decodes it without the
AV1 segfault; frames are extracted from the local file (ffmpeg never hits
the network, avoiding the static-build TLS crash).
- Same engine backs the MCP
generate_timestamped_pdf tool and the
build_zip.py package — one code path, three surfaces.
Exit codes
0 ok · 3 no video formats offered (IP blocked — add cookies/proxy) · 4 yt-dlp/ffmpeg missing · 5 other.
