What it does

Videocut Skills is a Claude Code Skill (think plugin) that automates the tedious cleanup of Chinese talking-head videos — the ums, repeated sentences, dead air, and verbal stumbles that editors usually hunt through manually. It transcribes via Volcano Engine ASR, runs semantic analysis through Claude, presents cuts in a browser for human approval, then fires FFmpeg to do the actual surgery.

The interesting bit

The author built this because Jianying’s “smart cut” only does pattern matching — it can’t tell when you’ve said something twice with slightly different wording, or when you started a sentence wrong and corrected yourself. The agent uses Claude to actually parse meaning, plus a custom dictionary file for tech terms like “MCP” that ASR routinely botches. There’s even a “self-update” skill where you tell it preferences (“keep some ums as transitions”) and it remembers.

Key highlights

• Detects 8 distinct problem types: silence >0.3s, repeated sentence openings, intra-sentence repetition, filler words, stutters, corrections, and more
• One documented run: 19-minute raw video → 608 detected issues → AI-assisted cutdown
• Browser-based review UI with click-to-play, double-click toggle, shift-drag multi-select
• Optional 2-pass sharpening export and Whisper-based subtitle burn-in with dictionary correction
• Rules live as markdown files the LLM reads — no retraining, just editing text

Caveats

• Requires Volcano Engine API key (Chinese cloud ASR), plus ~5GB of local models (FunASR + Whisper large-v3)
• Audio uploads to uguu.se by default — you’ll want to check that privacy tradeoff
• Port 8899 hardcoded for the review server; conflicts mean manual debugging

Verdict

Chinese-speaking developers who regularly edit their own talking-head content should try this — it genuinely reduces the cognitive load of finding “the good take.” Everyone else, or anyone allergic to cloud ASR dependencies, can wait for a more portable fork.