Youtube Publisher

youtube-publisher

The YouTube publishing pipeline. Takes raw video files on disk and ends with a fully-published series: titles in the right scheme, descriptions with the right CTA, custom on-brand thumbnails on every video, and a properly-ordered playlist. Idempotent state files make every step re-runnable so future series ship in minutes, not hours.

---

SETUP (run this once)

YouTube publishing needs your own Google Cloud project + OAuth. There is no shared key — this talks to your channel.

1. Google Cloud project → enable YouTube Data API v3.
2. OAuth Desktop client → download client_secrets.json (keep it

out of git — see .gitignore).

1. Verify the channel at https://www.youtube.com/verify (phone +

SMS). Required before custom thumbnails work, and before description links become clickable.

1. Fill the placeholders used in the schemes below: ``,

``.

1. The first real API call pops a browser for OAuth and caches the token

locally (token.json, auto-refreshed, gitignored).

This skill documents the pipeline (upload → render thumbnails → finalize → playlist) using the YouTube Data API v3. The two scripts referenced below (upload.py, finalize.py) are small, resumable, idempotent wrappers — set them up once in / (or have the agent build them from this spec) and reuse them per series.

---

When to use

Trigger on:

• "Upload these videos to YouTube"
• "Publish the [series name] videos"
• "Make a YT playlist for X"
• "Set custom thumbnails on these videos"
• "Update YT titles / descriptions"
• "Redo the thumbnails"
• User types /youtube-publisher

Skip for:

• Editing raw footage → /video-use
• Building motion-graphics launch videos → /launch-video
• Writing the YouTube script itself → /youtube-script
• Designing the thumbnail art → /youtube-thumbnail (interview) or

/graphics-designer (tutorial / solo). This skill sets the thumbnail; those make it.

---

Pipeline shape

/
├── upload.py            # resumable uploads, OAuth flow, idempotent (uploaded.json)
├── finalize.py          # thumbnails + titles + descriptions + playlist (finalize-state.json)
├── metadata.json        # per-video title/description/tags for the UPLOAD step
├── client_secrets.json  # OAuth Desktop client (gitignored, one-time setup)
├── token.json           # cached OAuth token (gitignored, auto-refreshed)
├── uploaded.json        # idempotency: file → {video_id, video_url, ...}
└── finalize-state.json  # idempotency: {thumbnails, titles, descriptions, playlist_id, playlist_items}

Keep the upload-time metadata (per-video tags/description footers) and the finalize-time meta (thumbnail design fields) separate. They reference the same video_id after upload.

Source videos large enough to need Git LFS (*.mp4, *.mov) should be tracked before committing — GitHub rejects files >100 MB without it.

---

Definition of done · HARD GATE

A series is NOT done until all six are true:

1. Source video files in place (Git LFS if >100 MB)
2. Each video uploaded to YouTube, video_id captured in uploaded.json
3. Each video's title set per the scheme below
4. Each video's description set per the scheme below (CTA + playlist link for series videos)
5. Each video has a custom thumbnail pushed
6. The series playlist exists and contains all videos in the intended order

Do not declare "done" until you can paste the playlist URL and every video's URL with the new title + thumbnail visible in Studio.

---

Title scheme

Topic first, series second. Topic wins for click-through; series tag is identification. Use the middot separator (no em dash, no hyphen, no colon — brand rule).

Series video:
  CLI Install · 
  MCP Server · 
  Connections · 
  ...

Standalone / non-series:
   Demo · 

For a standalone flagship video, the tagline can live in the title instead of a series tag.

---

Description scheme

Terse. CTA first. YouTube collapses descriptions after the first ~150 chars, so the link to `` goes in the very first line.

Series videos:

Try  free → 

Full  playlist → https://www.youtube.com/playlist?list=

Standalone:

Try  free → 

That's it. Don't pad with chapter timestamps, social links, hashtag walls. If the user wants more (timestamps from /youtube-description, related links), add them explicitly — never default to bloat.

Gotcha: Links in descriptions render as plain text until Google finishes a one-time identity verification on the channel (~24h). After that they're clickable automatically. Nothing to script — just warn the user once.

---

Thumbnail

Set the thumbnail produced by /youtube-thumbnail (interview/podcast) or /graphics-designer (tutorial/solo). One unified design per series — geometry stays identical, only the per-video kicker + title swap. Render at 1280×720, push via youtube.thumbnails().set() in the finalize step.

---

Workflow

1. Stage the videos

Drop the cut files into your series folder, numbered for order (NN-topic.mov). Track large files with Git LFS first if you version them:

git lfs track "media//*.mp4" "media//*.mov"
git add .gitattributes

Files >100 MB will be rejected by GitHub without LFS, so set this up first.

2. Configure upload metadata

Edit metadata.json. A _defaults block (category, privacy, tags, descriptionFooter) applies to every video; the videos[] array carries per-video title/description/tags + the file path.

Default privacy: unlisted. Always. Public-flip happens after the user reviews the live videos in Studio.

3. Upload

python3 /upload.py --dry-run --all   # preview
python3 /upload.py --only 0          # one as a smoke test
python3 /upload.py --all             # the rest

First real call pops a browser for OAuth on the channel-owner Google account. Token caches in token.json. Resumable uploads, 8 MB chunks, 500/502/503/504 auto-retry with a short sleep.

Quota: 1,600 units/upload × default 10k/day = 6 uploads/day. For 7+ videos either spread across a day boundary or request a quota increase via the YouTube API quota form (~1-3 weeks).

4. Render thumbnails

Produce them with /youtube-thumbnail or /graphics-designer. Required per video at finalize time: slug, video_id (from uploaded.json), kicker, title, yt_title. Sanity-check every PNG before pushing — the visual review here is the only thing between you and a bad thumbnail going live. Specifically check:

• Long single-word titles fit the column
• Kicker reads cleanly
• Logo + wordmark visible, NOT cropped
• The graphic is centered, not bleeding off

5. Finalize

python3 /finalize.py --dry-run    # preview
python3 /finalize.py              # ship

Does, in order:

1. Thumbnails — youtube.thumbnails().set() per video
2. Titles — youtube.videos().update() snippet
3. Descriptions — youtube.videos().update() snippet
4. Playlist — creates if missing, adds each video at position: idx

Flags:

• --only-thumbnails / --only-titles / --only-descriptions / --only-playlist
• --force — re-run actions already recorded in finalize-state.json
• --delay-between N — seconds between thumbnail uploads (avoids rate limit)
• --dry-run

6. Public flip

Last step the user controls. Either:

• Studio → each video → Visibility → Public (manual review)
• Or extend finalize.py with a --make-public flag that calls

videos.update with status.privacyStatus: "public" per video

Default to the manual review — the user wants to confirm the actual live thumbnail and metadata before publishing.

---

Known gotchas (each one bites at least once)

Custom thumbnails require channel verification. First attempt returns 403 The authenticated user doesn't have permissions to upload and set custom video thumbnails. Fix: user goes to https://www.youtube.com/verify, enters phone + SMS code. ~2 min. After that, thumbnails work.

Thumbnail upload rate limit. YouTube caps thumbnails.set per channel in a rolling window. A burst of 4-5 back-to-back returns 429 The user has uploaded too many thumbnails recently. The state file preserves what succeeded; pass --delay-between 120 (2 min between uploads) when re-running. Window is opaque but often clears in ~1-2h, sometimes resets at midnight PT. Fallback when blocked: open Finder + the Studio edit tabs and drag-drop manually — Studio's upload path isn't subject to the same limit.

open -R output/-00-.png
for vid in   ...; do open "https://studio.youtube.com/video/$vid/edit"; done

Identity verification for clickable links. First time you put URLs in any description, Google triggers a one-time channel identity verification (~24h). Until done, links render as plain text. Nothing to script.

Daily upload quota. 10k units/day default → 6 uploads/day. 7+ videos crosses a day boundary. Have the script catch quota errors and break the loop with a clear message — resume tomorrow.

Position-0 truthy check. if state["playlist_items"].get(vid): treats a position-0 entry as not-in-playlist because 0 is falsy. Use if vid in state["playlist_items"]: instead.

---

Don'ts

• Don't restart from scratch when state exists. uploaded.json and

finalize-state.json make the whole pipeline idempotent. Re-running any step is a no-op for what's already done. Use --force only when re-pushing because the artifact actually changed.

• Don't push thumbnails in a tight loop. The rate limit will bite.

Use --delay-between 120 for batches; fall back to Studio drag-drop if the API is already cooling off.

• Don't write rich descriptions. Terse CTA + playlist link is the

whole brief. Timestamps, social links, hashtag walls without the user asking is the kind of bloat that reads as AI slop.

• Don't auto-flip to public. Default is unlisted until the user

manually reviews each video in Studio. Public flip is an explicit ask.

• Don't change the thumbnail design without showing renders first.

Iterate visually, get a thumbs-up, then ship.

• Don't add an em dash anywhere. Brand rule. Title separator is the

middot (·).

• Don't bypass the channel verification step. Tell the user to verify

at youtube.com/verify before the first thumbnail upload, not after the first 403.

• Don't commit client_secrets.json or token.json. They're in

.gitignore already; double-check before any git add -A.

---

See also

• Thumbnail art: /youtube-thumbnail (interview) · /graphics-designer (tutorial/solo)
• Description + chapters: /youtube-description
• Script before filming: /youtube-script
• Raw footage edit: /video-use

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: timscheuerai
• Source: timscheuerai/content-vault
• License: MIT

Install and usage instructions live in the source repository linked above.