A video editor that actually understands your verbal stumbles

What it does

This is a set of Claude Code Skills that turns a raw talking-head video into a tight final cut by having Claude actually listen for meaning. It sends audio to Volcano Engine’s ASR for word-level timestamps, then uses Claude to flag repetitions, self-corrections, filler words, and silences. A local review web UI lets you approve or reject the cuts before FFmpeg trims the final video. It also handles subtitle generation with a custom dictionary to stop technical terms like Claude Code from being mangled.

The interesting bit

The clever part isn’t the FFmpeg plumbing; it’s the rule set encoded as markdown files under 用户习惯/ that Claude reads to learn editing preferences—like “delete the first take when I repeat a sentence” or “keep a little ‘um’ as a transition.” The project treats the LLM as an apprentice editor with a style guide, not just a script runner.

Key highlights

• Semantic editing: detects intra-sentence repetition, adjacent-sentence restarts, and negative corrections (“no wait, actually…”) rather than just waveform gaps.
• Human-in-the-loop: generates a browser-based review page with click-to-play segments; you confirm before anything is deleted.
• Custom terminology dictionary for subtitle burn-in, aimed at tech terms that generic ASR routinely botches.
• Self-updating rules: tell Claude your preferences in natural language and it writes them back to the rule files for future sessions.
• Adjustable silence threshold (default >0.3 s), unlike the fixed thresholds in consumer apps.

Caveats

• Requires a Volcano Engine API key and uploads audio to a third-party host (uguu.se) by default for transcription, which may raise data-privacy eyebrows.
• The review UI runs on a hardcoded local port (8899), so collisions are possible.
• It is explicitly designed for spoken-broadcast (口播) content; don’t expect it to handle vlogs, music, or multi-speaker scenes gracefully.

Verdict

If you regularly produce long, single-speaker tutorials or livestream replays and spend half your time manually scrubbing out verbal stumbles, this is a genuine time-saver. Everyone else can stick to their timeline.