wjs-uploading-video

Push finished videos to YouTube. Defaults are tuned for this user's workflow (王建硕 channel, China network with local proxy, 1080p horizontal recordings from Riverside / multicam edits).

When to use

• User has one or more finished .mp4 files and wants them on YouTube
• User points to a final/ directory with multiple segments and an UPLOAD_META.md
• User wants a specific privacy / playlist / scheduled publish

Don't use for:
- 微信视频号 upload (no public API; user uploads manually via web)
- 抖音 / 小红书 / B 站 (different APIs, not yet implemented here)
- YouTube Shorts variants from horizontal source (use wjs-reframing-video first to produce the 9:16 cut, then upload that via this skill)

Prerequisites (one-time per machine)

1. Google Cloud OAuth client: ~/.config/youtube/credentials.json must exist. See references/credentials-setup.md for the 5-minute setup if missing.
2. Python deps: pip3 install google-auth-oauthlib google-api-python-client requests (only google-auth + requests are strictly needed at upload time, but the OAuth-lib pulls them).
3. First-ever upload opens a browser for Google consent and writes ~/.config/youtube/token.json. Subsequent runs reuse it silently.

How it works (and why it's not the stock youtube-uploader)

YouTube's resumable upload protocol issues a Location: URL after the metadata POST, then accepts the bytes in chunked PUT requests. The stock google-api-python-client runs this over httplib2, which under this user's local SOCKS+HTTP proxy stack throws [Errno 65] No route to host or socket.timeout on those follow-up PUTs and stalls indefinitely.

This skill bypasses httplib2: it does OAuth via google-auth, then drives the resumable upload manually with requests, passing the proxy explicitly. 8 MB chunks (not the stock 256 KB) — fewer round-trips through the proxy. Exponential-backoff retry on socket.timeout / ConnectionError / 5xx.

If you're tempted to "just call the YouTube API client directly," don't — it'll fail in this environment.

ModuleNotFoundError: No module named 'google.oauth2' — Homebrew 升级 Python 大版本后 site-packages 会丢（2026-08 从 3.13→3.14 实际发生过）。别重新 pip install，直接用 uvx 带依赖跑（已验证）：

uvx --with google-auth --with requests python \
  ~/.claude/skills/wjs-uploading-video/scripts/upload_youtube.py --dir "/path/to/final"

Usage

Batch upload a final/ directory

python3 ~/.claude/skills/wjs-uploading-video/scripts/upload_youtube.py \
  --dir "/path/to/final" \
  --meta "/path/to/final/UPLOAD_META.md"

The script:
1. Reads UPLOAD_META.md and pairs each ## NN · filename.mp4 block to a video file in --dir
2. Skips videos already in --results-file (default <dir>/.youtube_upload_results.json) — safe to re-run after failures
3. Uploads sequentially with progress every 8 MB
4. Writes the final URL list to --results-file

Single file

python3 ~/.claude/skills/wjs-uploading-video/scripts/upload_youtube.py \
  --video /path/to/clip.mp4 \
  --title "My Title" \
  --description "Body text" \
  --tags "tag1,tag2,tag3"

Common overrides

UPLOAD_META.md format

The parser expects the user's standard structure:

## 01 · segment_01_no-bugs.mp4

**短标题**
代码没有错误,只有意图不一致

**视频描述**
AI 时代屎山的重新定义...

—— 王建硕 × 任鑫《...》第 1 集

#王建硕 #AI编程 #ClaudeCode

---

Mapping:
- ## NN · <filename> → which video this block describes
- **短标题** (or **Title**) block → YouTube title, verbatim. Short titles work but consider that YouTube allows up to 100 chars — if you want a richer title with series name, write it that way in **短标题**
- **视频描述** (or **Description**) block → YouTube description, verbatim, with the #tag hashtags retained at the bottom
- All #word tokens in the 视频描述 → comma-separated YouTube tags (each # is stripped; the user's channel name 王建硕 is auto-prepended per global instructions)

Filename in the heading must match an actual file in --dir. If a file exists but has no meta block, the script errors loudly — pass --allow-missing-meta to upload it with --title <basename> and empty description.

Sensible defaults this skill bakes in

• Privacy = public: videos go live immediately and appear in subscriber feeds + search. Override per call with --privacy unlisted (link-only) or --privacy private (owner-only) when you want to review first
• Category = 28 (Science & Tech): matches 王建硕 channel's main content; override with --category per upload
• selfDeclaredMadeForKids = false: YouTube requires this; the user's content is adult-targeted
• Chunk size = 8 MB: validated working size through this user's local proxy (256 KB stalled)
• Skip already-uploaded: results file is the source of truth; deleting it forces re-upload

Channel-name CTA rule

If you write description footers, signatures, or "subscribe to me" lines into a video's metadata, use 王建硕 (the user's channel name). Don't put a guest's name there — guests like 任鑫 belong inside the description body when they're the conversation partner, never in the channel-CTA slot.

Troubleshooting

After uploading

• Results saved to --results-file (JSON: file, title, video id, URL)
• Echo the URL list back to the user in a markdown table
• Videos are public by default — they're live the moment the upload completes. If the user wanted a review buffer, mention they can re-upload with --privacy unlisted (or flip back to unlisted in YouTube Studio)
• If they mentioned a 合集 / series, offer to create a YouTube playlist and batch-add the new uploads (the user must give a playlist ID or let you create one)