Loom Vision

Loom Vision

Use this skill whenever:

• The user pastes a Loom share URL (https://www.loom.com/share/<32-hex-id> or similar).
• The user asks you to "watch", "review", "summarize", or "explain" a Loom video.
• The user references a Loom video they recorded and asks a question that needs visual context (e.g. "what was the error on screen at minute 3", "did I click the right button", "what's the diff in that PR walkthrough").

The skill exists because Loom's auto-transcript captures only what was said aloud - not what was on screen. For code walkthroughs, design reviews, and bug repros, the screen contents are often more important than the audio.

How to invoke

Run the processor script with the Loom share URL. It downloads the mp4, samples frames, and prints the output directory path.

bash {baseDir}/process-loom.sh ""

The script prints the output directory path to stdout. Inside that directory:

• video.mp4 - the original video (kept in case you want to re-process at a different frame interval)
• transcript.vtt - Loom's auto-transcript as WebVTT with timestamps (read this one)
• transcript.txt - the same transcript as plaintext, no timestamps
• transcript.json - the raw loom-dl transcript JSON, kept verbatim
• frame_001.jpg, frame_002.jpg, ... - sampled frames at 1 frame per N seconds (default 5)

Note: loom-dl writes its transcript as .transcript.json, not VTT. The script converts that JSON into transcript.vtt + transcript.txt for you, so the "read the transcript" step below always has a file to read.

What to do with the output

Read the transcript first to anchor timing, then browse the frames that fall in the relevant time window. For a 10-minute video at 5-second sampling that's 120 frames - review them in batches or jump to the specific window the user asked about.

Frame N corresponds to timestamp (N - 1) * FRAME_INTERVAL_SECONDS from the start of the video. Default interval is 5 seconds, so frame_007.jpg is at the 30-second mark.

Working with the output

Common patterns:

• Summarize the whole video. Read the full transcript, then sample every 6th frame (one per 30 seconds) for the visual narrative. Combine.
• Answer a specific question. Search the transcript for the relevant keyword, find the timestamp, look at the frame(s) within plus or minus 10 seconds.
• Spot what's missing from the transcript. Errors, popup dialogs, UI states, and on-screen code rarely make it into the spoken audio - go directly to the frames.

Configuration

All knobs are environment variables, set before invoking the script:

• FRAME_INTERVAL_SECONDS - default 5. Lower for densely visual content (animations, fast UI changes), higher for slow walkthroughs.
• FRAME_MAX_WIDTH_PX - default 1280. Bump up for tiny-font code review videos.
• FRAME_QUALITY - default 3 (ffmpeg -q:v scale, 1-31, lower = better).
• OUTPUT_ROOT - default ${TMPDIR:-/tmp}/loom-vision.

Dependencies

Three system CLIs are required (declared in the install specs above so OpenClaw can offer to install them):

• node - runs the transcript JSON to VTT conversion. It is also required by loom-dl, so if loom-dl works, node is already present. Install with brew install node if missing.
• loom-dl - install with npm install -g loom-dl (Node CLI, not on Homebrew)
• ffmpeg - install with brew install ffmpeg (macOS) or your distro's package manager (Linux). Supplies ffprobe, used to bound the final transcript cue.

All installs are idempotent.

If your ffmpeg install lacks ffprobe (notably the ffmpeg-static npm module), the script degrades gracefully: it estimates the end time of the last transcript cue instead of reading the true video duration. Frames, transcript text, and all earlier cue timings are unaffected.

Why this exists

Most Loom-style integrations treat the video as audio-with-pictures: they ingest the transcript and stop. That loses the entire visual half of a walkthrough - IDE state, error popups, on-screen code, design mockups, the actual change in a PR review.

Sampling at 1 frame per 5 seconds is the sweet spot for code/UI content: dense enough to catch transient errors and UI transitions, sparse enough that a 10-minute video yields 120 frames (manageable batch size for multimodal review). The transcript provides the audio narrative; the frames provide everything else.