Automate a Podcast Pipeline and VPS Hardening with wlzh Skills

Turning a YouTube video into a Chinese podcast episode normally means juggling a downloader, a transcriber, an audio editor, and a publishing tool. The 606-star wlzh/skills collection packages that chain — and a stack of other daily grunt work — as Claude Code skills you invoke with a slash.

Why This Skill Matters

The repo is a practical, Chinese-language collection of Claude Code skills: media conversion, scraping, publishing, expense reports, and server ops. The README catalogs skills such as voice-changer (RVC AI voice conversion with presets), audiocut-keyword (FunASR Paraformer transcription with character-level timestamps, then FFmpeg cutting), text-to-speech (Edge TTS with 18+ Chinese voices and SSML emotion marks), youtube-publisher (metadata, thumbnails, SRT/VTT subtitles, dry-run preview), wlzh-invoice-scanner, and vps-security-hardening.

Two caveats from checking the repo against its README on September 15, 2026: the README lists 16 skills but the repository carries 15 skill directories — the youtube-to-xiaoyuzhou entry's directory was not present as of that date. Two entries also map to differently named directories: the README's youtube-tracker corresponds to wlzh-youtube-tracker, and x-fetcher to x-fetcher-skill. The composition story below is the README's documented example, not a directory you can install today.

Installation

The README documents usage, not a copy procedure — its usage lines assume the skill directories already sit under ~/.claude/skills/, and it refers you to each skill's own README for dependencies. Once a skill is in place, the two documented invocation styles are:

# direct script call
python3 ~/.claude/skills/<skill-name>/scripts/script.py [arguments]

# or from a Claude Code session
/<skill-name> [arguments]

Real Workflow: Cut, Revoice, and Republish a YouTube Episode

The README's composition example chains the podcast pipeline in one line:

youtube-to-xiaoyuzhou <url> --filter-keywords --change-voice female_3

The README's youtube-to-xiaoyuzhou entry documents the pipeline's own stages: automatic YouTube download, AI cover generation (via image-generator), keyword filtering (via audiocut-keyword), voice conversion (via voice-changer), scheduled publishing, and cookie-first login. The same pieces exist as individually documented skills: video-downloader fetches videos with quality options (best, 1080p down to 360p) or audio-only MP3, and audiocut-keyword on its own transcribes with character-level timestamps and removes segments matching your keyword config. A shorter chain is also documented: text-to-speech script.txt --voice female_3 turns a script into speech and can post-process it through the voice changer.

Real Workflow: Harden a Fresh VPS

vps-security-hardening is the one skill in the repo whose README entry documents a full command-line interface — a seven-step hardening pass over SSH:

./scripts/harden-vps.sh \
  --ip <VPS_IP> \
  --root-pass <ROOT_PASSWORD> \
  --user <NEW_USER> \
  --user-pass <NEW_USER_PASSWORD> \
  --port <SSH_PORT>

The seven steps it runs: create a sudo user and disable root password login, change the SSH port (with Ubuntu version detection for 22.10/23.x versus 24.04+), install and configure Fail2ban, support SSH key login, optional SSH login notifications, configure UFW, and a Docker safety reminder.

Tips

• voice-changer does real timbre conversion with RVC AI models — not just pitch adjustment — with presets such as 中文御姐, 七妹, 苏苏, and AZI, and it is explicitly designed to be callable from other skills, which is what makes chaining it into the podcast pipeline possible.
• youtube-publisher supports a dry-run mode that previews the configured settings without uploading — check your setup there before a real publish.
• Several skills wrap upstream projects and say so: video-downloader is sourced from ComposioHQ/awesome-claude-skills, x-fetcher is based on Jane-xiaoer/x-fetcher, and wespy-fetcher on tianchangNorth/WeSpy.
• wlzh-invoice-scanner (v3.6.0) sorts invoices into five categories, verifies amounts with Python code rather than LLM mental math, and dedupes by invoice number.

When Not to Use This

The collection is Chinese-first: TTS voices, voice presets, and the invoice formats target Chinese workflows, and the README is in Chinese. The heavier media skills assume local dependencies (FFmpeg, FunASR, RVC models) per each skill's README. And since the repo documents no install command, you are copying directories by hand — fine for one machine, tedious across a team.

See the leaderboard for more skills.