yt-dlp-cli

yt-dlp. Use when the user wants to download, extract audio, pull subtitles, clip, or archive a YouTube or other streaming-site URL, or when yt-dlp hits 403, bot-check, or only-360p. Not for local-file edit/transcode, watch/summarize without saving, search, or generic downloads.

yt-dlp-cli - agent workflow

Result: After a user asks to download, extract, clip, subtitle, or archive media from a yt-dlp-supported URL, the agent has either produced the correct file(s) on disk and reported path/size/identity, or stopped on an explicit blocker (missing binary, auth required, rights unclear, playlist too large, YouTube bot wall).

Next consumer: The user (file on disk) and any later skill that transcribes/edits that file.

Done: File exists at the stated path, matches the requested job (video vs audio vs clip vs playlist range), and the agent reported what it did and how it verified. Or a named blocker with the next safe action.

Intent (changes the approach): Agents can already invoke yt-dlp <url> and they do it badly. This skill exists to close those failure modes, not to teach the CLI. Do not write a flag encyclopedia. Do not hide yt-dlp behind a Python wrapper.

Autonomy envelope

Invoking this skill authorizes local inspection, downloading the requested streaming-site URL, writing the resulting file(s) under the chosen destination, and running required local ffmpeg postprocessing for the requested media shape. It does not authorize downloading an unbounded playlist/channel, reading browser cookies or auth files, bypassing unclear rights, or supplying user-only inputs. Stop for confirmation only when rights are unclear, scope expands to a full playlist/channel without an agreed range/archive, cookies/auth are needed, or the next input can only come from the user.

Quote every URL in shell commands. Treat a watch URL as one video unless the user explicitly asked for the playlist; keep --no-playlist on single-video jobs. Check ffmpeg -version before merged video/audio, audio extraction or conversion, subtitle embedding, clips, or SponsorBlock postprocessing; missing ffmpeg is a blocker for those shapes.

Default download

For a single requested video, run this command first:

bash

yt-dlp --no-playlist -P "<dir>" -o "%(title).200B [%(id)s].%(ext)s" --merge-output-format mp4 --print after_move:filepath "<url>"

Use the path printed by --print after_move:filepath as the authoritative handoff path. Verify that file exists and is non-empty; for video, verify duration and resolution with ffprobe or equivalent.

Deltas from the default

• Height cap: add -f "bv*[height<=1080]+ba/b[height<=1080]" and change only the numeric cap. Use yt-dlp --no-playlist -F "<url>" first only when the user requested a specific format ID or the first run returns the wrong shape; never invent format IDs.
• Audio-only: replace --merge-output-format mp4 with -x; add --audio-format <format> only when the user requested a container/codec such as mp3 or m4a.
• Subtitles: for subtitle-only jobs, replace media download/postprocessing with --skip-download --write-subs --sub-langs "<langs>". For video plus subtitles, keep the default download and add --write-subs --sub-langs "<langs>". Add --write-auto-subs only when generated captions are acceptable, and --embed-subs only when embedded subtitles are requested.
• Output layout: change only -P "<dir>" and -o "<template>". Keep %(id)s in the template unless the user requested a different stable naming scheme.
• SponsorBlock: add --sponsorblock-remove <categories> only for categories the user requested or approved.

Pinned fragile routes

For playlist or channel archive work, first confirm the range or archive scope. Then use an archive file so reruns skip completed entries:

bash

yt-dlp -I "1-10" --download-archive "<dir>/download-archive.txt" -P "<dir>" -o "%(playlist_title).120B/%(playlist_index)s - %(title).200B [%(id)s].%(ext)s" --print after_move:filepath "<playlist-or-channel-url>"

Keep -I for item ranges; remove it only after the user confirmed a full archive. The archive file is part of the handoff because it prevents duplicate downloads on rerun.

For clips or time sections, use --download-sections and verify the resulting duration against the requested span:

bash

yt-dlp --no-playlist --download-sections "*00:01:30-00:02:10" -P "<dir>" -o "%(title).200B [%(id)s] clip.%(ext)s" --merge-output-format mp4 --print after_move:filepath "<url>"

Change only the timestamps unless the user requested a different section.

Ordered hatch

Run the pinned default or fragile route first. If it exits non-zero, returns the wrong media shape, reports a bot/auth/version mismatch, or shows only 360p where higher formats should exist, take this order:

1. Preserve the command, exit status, and stderr.
2. Update yt-dlp, verify yt-dlp --version, then retry a low-cost inspection with yt-dlp --no-playlist -F "<url>" or yt-dlp --no-playlist --print ... "<url>".
3. For wrong shape or site-specific syntax drift, consult live yt-dlp --help / yt-dlp -h and adjust the smallest delta needed.
4. For YouTube 403, "confirm you're not a bot", missing higher qualities, nsig errors, PO-token paths, or header-sensitive HLS failures, read references/youtube-auth.md in full before retrying.
5. Stop with a blocker when rights, scope, auth, user-only input, or a missing binary prevents safe completion.

Report

The report may omit secondary detail, but it must preserve command, exit status, output path and size, media identity when available, verification performed, and stderr or blocker. For video include duration/resolution; for audio include audio-only confirmation; for playlists/channels include item count and archive-file path.