loom-analyze (Claude Code skill)

Turn a Loom share URL into a transcript (and optional keyframes) locally — no API key, no third-party MCP.

Install

Copy this entire folder to ~/.claude/skills/loom-analyze/ on the target machine.
Run the one-time setup:
```
bash ~/.claude/skills/loom-analyze/setup.sh
```
Installs Homebrew packages yt-dlp + ffmpeg and pip package openai-whisper.
Restart Claude Code. The skill is auto-discovered from the SKILL.md frontmatter.

Use

Inside Claude Code, paste any https://www.loom.com/share/... URL or say:

“transcribe this loom "
“analyze this loom "
“/loom-analyze "

Claude will call the shell wrapper and inline the transcript.

Direct CLI use:

~/.claude/skills/loom-analyze/loom-analyze.sh <URL>                 # transcript only
~/.claude/skills/loom-analyze/loom-analyze.sh <URL> --frames 10     # +PNG every 10s
~/.claude/skills/loom-analyze/loom-analyze.sh <URL> --model small   # better accuracy
~/.claude/skills/loom-analyze/loom-analyze.sh <URL> --keep          # don't delete MP3/MP4

Output: ~/Downloads/loom-transcripts/<id>/

audio.txt — transcript
frames/frame-NNNN.png — keyframes (when --frames used)

Requirements

macOS (or Linux with Homebrew). Windows + WSL works the same.
Python 3 (for whisper).
~140 MB disk for the base whisper model (downloaded on first run); small ~460 MB; medium ~1.5 GB; large ~3 GB.

Troubleshooting

“Missing dep” error: re-run setup.sh.
whisper not found after install: ensure ~/Library/Python/3.x/bin (or wherever pip user-installs binaries) is on PATH.
Loom URL not recognised: must be loom.com/share/<id>. Embedded/private links won’t work without auth.

What it does NOT do

No streaming / real-time. Whole-video transcription only.
No speaker diarization (single transcript stream).
No upload to any third-party service. All processing is local.