watch

/watch: claude-video-plus

Same answer quality. Half the tokens. Faster than the default.

Ask a video a question and /watch fetches only the evidence that answers it. In a sealed, receipt-gated benchmark on five videos it had never seen, evidence mode matched the original's blind-judged answer quality (8.83 vs 8.80) while sending the AI 56% less material, and it ran ~14 seconds faster end-to-end than the default mode. Zero risk to adopt: it only runs when you ask a question, and on any problem (no captions, local file, short video, any error) it runs the original pipeline instead.

[!NOTE] This is a fork of Brad Bonanno's bradautomates/claude-video, not the original repository. Upstream history, MIT license, and authorship are preserved; see Gratitude.

Website: abe238.github.io/claude-video-plus · Guide: Ask Claude to watch a video · All benchmark data: docs/benchmarks/ · Changelog: CHANGELOG.md · Status: stable, deterministic suite on a hosted macOS/Linux matrix, isolated install lifecycle (status page)

Install

Claude Code (recommended, auto-updates via marketplace):

/plugin marketplace add abe238/claude-video-plus
/plugin install watch@claude-video-plus

Codex, Cursor, Copilot, Gemini CLI, or any of 50+ Agent Skills hosts:

npx skills add abe238/claude-video-plus -g

Zero config to start: yt-dlp and ffmpeg install on first run via brew on macOS (Linux/Windows print exact commands). Captions cover most public videos for free.

No API key required, and none is asked for. When a video has no captions, transcription falls to the local backends first — a loopback STT server, YAP on macOS, or the openai-whisper CLI on any platform — and audio never leaves your machine. Cloud Whisper exists but is doubly opt-in: a key alone does nothing without --allow-remote-transcription.

The numbers

The sealed confirmatory run: five videos the pipeline had never seen, ten questions and answer keys frozen by two independent annotators before any pipeline ran, three blind judges per video, and a one-time cryptographic receipt so the run cannot be quietly repeated. Raw data.

Result	Original	Evidence mode
Blind answer quality (1–10)	8.80	8.83, same quality
Required facts covered	3.57	3.67, same coverage
Material sent to the AI	100%	44% (56% less)
Cold end-to-end wall time	32–34 s	18–19 s (measured)

Development data tells the same story with more spread: 60–88% savings on targeted questions, wins concentrated on videos over 9 minutes, and every loss published unedited (deep-dive, battery).

How evidence mode works

/watch <url> --detail evidence --question "what's actually new, skip the hype?"

The original samples the whole timeline (50–100 frames plus the full transcript) no matter what you asked: ~50k tokens on a 40-minute video. Evidence mode retrieves instead: whole topical chapters (YouTube chapters, pause-gap fallback), a numeric guard that rescues pricing/benchmark/spec lines from anywhere in the video with a frame at each (numbers live on-screen), frames at chapter starts and "as you can see" moments, and a token budget. Every selection lands in a manifest with its timestamp, reason, and score, so you can audit exactly what the model saw and why.