Loom Analyze

Local pipeline (no API key, no third-party MCP) that turns a Loom share URL into:

1. A plain-text transcript via whisper.
2. Optional keyframe PNGs for visual analysis when the user asks about UI / screen content.

When to invoke

• User pastes a https://www.loom.com/share/... URL.
• User says “transcribe”, “analyze”, or “watch this loom”.
• /loom-analyze <url> is typed.

How to run

The shell wrapper lives next to this file: loom-analyze.sh. Always call it through Bash.

~/.claude/skills/loom-analyze/loom-analyze.sh <URL> [--model M] [--frames N] [--keep]

Flags:

• --model base|small|medium|large — whisper accuracy/speed tradeoff. Default base (~140 MB download first time).
• --frames N — also extract one PNG every N seconds into frames/. Skip this for transcript-only (faster).
• --keep — keep MP4/MP3 after transcribing. Default deletes them.

Output goes to ~/Downloads/loom-transcripts/<id>/. The transcript is printed to stdout, so the agent gets it in the tool result.

Decision rules for the agent

• Transcript-only is the fast default. Don’t request frames unless the user asks about visuals, UI, screen content, or a demo walkthrough.
• Long videos (>10 min): stick to --model base. --model small or larger only when the user specifically complains about transcript accuracy.
• When frames are extracted: read the PNGs from ~/Downloads/loom-transcripts/<id>/frames/ via the Read tool when the user asks what’s on screen at a given time. Don’t bulk-read every frame; pick the ones that match the question.
• First-time setup: if the script exits with “Missing dep”, run ~/.claude/skills/loom-analyze/setup.sh once, then retry.

Setup (per-machine, one-time)

bash ~/.claude/skills/loom-analyze/setup.sh

Installs yt-dlp, ffmpeg, openai-whisper via Homebrew + pip. Idempotent.