alexeygrigorev/.agents/youtube-manager

YouTube manager

Channel writes go through ~/git/youtube-manager-agent and the YouTube Data API. That repo is the source of truth for CLI flags, chopping, and manifests — read its README (and docs/chopping.md when cutting recordings). This skill is the agent workflow: auth, then the smallest API change that does the job.

Do not use this for transcripts/captions. Use fetch-youtube.

Setup

Always run from the agent repo with uv:

cd ~/git/youtube-manager-agent
uv run python -m auth.reauth --check

Expect status: VALID and authorized as: DataTalksClub. Credentials live in the gitignored .youtube/ folder (client_secret.json, token.json). Never commit them, print them, or copy them into the skill.

If --check reports DEAD / invalid_grant, stop. Ask the user to re-consent on a machine with a browser:

cd ~/git/youtube-manager-agent
uv run python -m auth.reauth

Then they can drop the fresh token.json at .youtube/token.json (or bash bin/push_token.sh to a remote checkout). Do not start run_local_server headless.

Course video links

DataTalks.Club course videos must link the current GitHub repo, not the old bookcamp tree:

• Lesson / notebook / repo URLs: https://github.com/DataTalksClub/<course-repo>/...
• Machine Learning Zoomcamp files live at repo root on main, not under

course-zoomcamp/.

• Do not restore stubs in alexeygrigorev/mlbookcamp-code to un-404 an old

YouTube URL. Rewrite the description.

Typical ML Zoomcamp rewrite:

Verify rewritten GitHub URLs return 200 before telling the user it is done.

ML Zoomcamp playlist: PL3MmuxUbc_hIhxl5Ji8t4O6lPAOpHaCLR.

Commands

cd ~/git/youtube-manager-agent
 
uv run python -m playlist.list_playlist PLxxxxxxxx
uv run python -m playlist.add_to_playlist --playlist PLxxxxxxxx [--position N] [--before-title "…"] [--dry-run] <id-or-url>…
 
uv run python -m video.rename --manifest manifests/<name>.json --dry-run
uv run python -m video.rename --manifest manifests/<name>.json
 
uv run python -m video.add_chapters --manifest manifests/<name>.json --chapters-dir work/chapters --dry-run

Playlist positions are 0-based. add_to_playlist is idempotent: without --position it skips videos already in the playlist; with --position it moves them.

Publish chopped clips with manual Studio upload + video.rename, not video.upload. API videos.insert stays private until the Google Cloud project is audited.

Chopping pipeline (download → spec → bin/chop.sh → captions/chapters): see docs/chopping.md. Pull source transcripts with fetch-youtube.

Read or rewrite descriptions

There is no description-only CLI. Use videos.list / videos.update via auth.auth.get_service. videos.update with part=snippet must send title, categoryId, and description. Keep existing tags and defaultLanguage when present.

Quota: videos.update ≈ 50 units, videos.insert ≈ 1,600, default daily cap 10,000. Dry-run bulk edits (print id + title, count matches) and apply only after the user would accept the replacement.

from auth.auth import get_service
yt = get_service()
vid = "xxxxxxxxxxx"
sn = yt.videos().list(part="snippet", id=vid).execute()["items"][0]["snippet"]
new_desc = sn["description"].replace(OLD, NEW)
body = {
    "id": vid,
    "snippet": {
        "title": sn["title"],
        "description": new_desc,
        "categoryId": sn.get("categoryId", "27"),
    },
}
if sn.get("tags"):
    body["snippet"]["tags"] = sn["tags"]
if sn.get("defaultLanguage"):
    body["snippet"]["defaultLanguage"] = sn["defaultLanguage"]
yt.videos().update(part="snippet", body=body).execute()

For a playlist, page playlistItems.list, then videos.list in batches of 50.

After a write, fetch the snippet again and confirm the new text (and that old prefixes are gone).