video-summary

name	video-summary
description	Summarize a video (URL or local path) by pulling captions first, falling back to audio-only transcription, and optionally extracting frames for visual summaries. Use when the user pastes a YouTube/Vimeo/X/etc URL, points at a local video file, or asks to summarize, watch, or extract takeaways from a video. Triggers on "summarize this video", "what's in this video", "watch this", URLs ending in /watch, /shorts, common video extensions (.mp4/.mov/.mkv/.webm).
allowed-tools	["Bash(bash ~/.claude/skills/video-summary/bin/run.sh *)","Bash(bash ~/.claude/skills/video-summary/setup.sh *)"]

video-summary

Local-first video summarizer. For URLs, yt-dlp first pulls metadata and captions without media download. If captions are missing or unusable, it downloads audio only and transcribes on-device with whisperkit-cli. Full video download and frame extraction happen only with --with-frames.

No external API, no Python venv. macOS / Apple Silicon only.

Step 0: setup check (silent on success)

Skip if you already ran the script once this session.

test -d ~/.cache/whisperkit/models/argmaxinc/whisperkit-coreml/openai_whisper-large-v3-v20240930 \
  && command -v ffmpeg >/dev/null \
  && command -v yt-dlp >/dev/null \
  && command -v whisperkit-cli >/dev/null \
  && command -v jq >/dev/null \
  && echo OK

If output is OK, proceed. Otherwise tell the user:

Setup is incomplete. Run bash ~/.claude/skills/video-summary/setup.sh once. First-time model download is ~626MB (1-10 min depending on connection).

Do not run setup.sh yourself. The user reviews and confirms.

Step 1: parse the user input

Separate the video source from any specific question.

• summarize https://youtu.be/abc -> source = URL, question = none
• https://youtu.be/abc what's the main argument? -> source = URL, question = "what's the main argument?"
• ~/Movies/talk.mp4 -> source = path, question = none

Step 2: run the skill

bash ~/.claude/skills/video-summary/bin/run.sh "<source>"

run.sh orchestrates watch.sh (metadata/captions -> audio fallback -> optional frames -> transcript) and summarize.sh (haiku CLI subprocess that writes the structured summary). Frame image tokens are used only when --with-frames is passed.

Optional flags forwarded to watch.sh:

• --with-frames download full video and extract frames; off by default
• --max-frames N soft cap on frames; default 100. Applies only with --with-frames
• --height H caps full-video download resolution and frame scale (default 480p). Applies only with --with-frames
• --start T / --end T focus on a section
• --fps F override auto-fps (cap 2)
• --language LANG force whisperkit language (en, pt, es, ...)
• --refresh wipe the whole work dir (download, frames, transcript, report, summary)
• --refresh-download re-download (cascades to frames + transcript + report + summary)
• --refresh-frames re-extract frames (cascades to report + summary)
• --refresh-transcript re-transcribe (cascades to report + summary)

Unreadable frames

Frames are disabled by default. When --with-frames is used, frames default to 480p. Do NOT lower this default - past runs at 240p caused VLM OCR failures (couldn't read code/slides/captions).

The summarizer subagent flags unreadable frames in its output via a FRAMES_UNREADABLE: line, and run.sh echoes the same signal to stderr:

[video-summary] FRAMES_UNREADABLE: <path1>, <path2>, ...
[video-summary] re-run with --refresh --height 720 (or 1080) to escalate

When you see this signal, do NOT silently accept the summary. Tell the user which frames were unreadable and offer to re-run at higher resolution:

bash ~/.claude/skills/video-summary/bin/run.sh "<source>" --with-frames --refresh-frames --height 720

Higher tiers: 720 -> 1080. Network cost roughly 2-3x per step. The cached video is invalidated by --refresh and re-downloaded at the new cap.

Local flag handled by run.sh:

• --refresh-summary ignore the summary cache only (re-run the haiku pass; useful after editing video-summary-prompt.md)

Resume points

Pipeline is composable. Each stage is gated by its own artifact under ~/.cache/video-summary/<id>/:

Stage	Artifact	Refresh flag
download	download.json + download/	--refresh-download
frames	frames.tsv + frames/	--refresh-frames
transcribe	transcript.json	--refresh-transcript
summarize	summaries/summary-*.md	--refresh-summary

Cascade rule: each --refresh-* wipes its own stage and everything downstream. Re-running on the same source replays cached stages instantly. Delete any artifact directly (e.g. rm transcript.json) to redo that stage without passing a flag.

Decision tree (pick the matching flag for the user's intent):

• "Re-summarize, transcript is fine" -> --refresh-summary
• You edited video-summary-prompt.md -> --refresh-summary
• "Transcript is bad, re-transcribe and re-summarize" -> --refresh-transcript
• "Need visual context" -> --with-frames
• "Frames are unreadable" -> --with-frames --refresh-frames --height 720
• "Re-download at higher resolution" -> --with-frames --refresh-download --height 720
• "Wipe everything" -> --refresh

Stdout of run.sh is the clean structured summary (TLDR / Verdict / Summary / Key moments / Caveats / Pacing) plus ## Processing strategy and ## Run metrics blocks. The FRAMES_UNREADABLE: and WORK_DIR: subagent lines are stripped before output.

The work dir path is in the footer of pre-summary-context.md earlier in the pipeline. To recover it for follow-ups, parse Work dir: from that file or use the cache root ~/.cache/video-summary/<id>/ where <id> is derived from the source URL/path.

The legacy alternative if claude CLI is unavailable: use the Agent tool with subagent_type: general-purpose and model: haiku, and pass the contents of video-summary-prompt.md plus pre-summary-context.md as the prompt body.

Artifacts

After a complete run, ~/.cache/video-summary/<id>/ contains reusable stages for both text-only and --with-frames runs:

• download/video.mp4 - source video. Present only with --with-frames.
• download/video.info.json - full yt-dlp metadata dump.
• download/video.<lang>.vtt - manual or auto captions, when available.
• download/audio.<ext> - audio-only fallback when captions are missing or unusable.
• download/thumbnail.<ext> - thumbnail image, when yt-dlp provides one.
• download.json - manifest with paths + title/channel/chapters/description.
• frames/frame_NNNN.jpg - extracted frames at the chosen height/fps. Present only with --with-frames.
• frames.tsv - <seconds>\t<absolute path> per frame. Use this to map a frame to its timestamp - do NOT estimate from frame_NNNN indices.
• audio.wav - 16kHz mono PCM. Only present for whisper fallback input.
• transcript.json - [{start, end, text}]. Canonical transcript for the video. Read directly for transcript follow-ups.
• pre-summary-context.md - assembled markdown consumed by the haiku subagent. It includes frame paths only with --with-frames.
• summaries/summary-<signature>.md - final structured output for each option set.
• summary.md - copy of the latest emitted summary for convenience.
• download.metrics.json / frames.metrics.json / transcribe.metrics.json - per-stage {seconds, cached, ...}.
• metrics.json - merged per-stage view for run.sh.
• claude_metrics.json - haiku subprocess cost / tokens / duration.
• frames_unreadable.txt - sidecar; present only when the subagent flagged unreadable frames on the last run.

Step 3: handle follow-ups

The user may ask follow-ups about the same video:

• "What was on screen at 2:15?" Re-run with --with-frames if frames are absent. If frames exist, read the relevant <work_dir>/frames/frame_NNNN.jpg directly. Do not re-dispatch the subagent for one frame.
• "Re-summarize focused on the demo section" Re-run run.sh with --start/--end; pipeline cache hit on the same source reuses cached media.
• "Check the transcript around MM:SS" Read <work_dir>/transcript.json directly.

If the user asks about a different video, run run.sh on the new source. Working dirs are independent per video ID.

Step 4: cleanup

Cache lives at ~/.cache/video-summary/. Do not delete entries proactively - they are the cache. The user can rm -rf ~/.cache/video-summary/<id> themselves when they want to drop one.

Failure modes

• Setup incomplete -> point to setup.sh. Do not run it yourself.
• yt-dlp fails on a URL -> the script prints the install age. If

  14 days, suggest brew upgrade yt-dlp and ask the user to retry.

• whisperkit-cli fails -> hard fail. Surface the error verbatim. Do not fall back to frames-only.
• No captions and no audio track -> ffmpeg audio extract fails; same hard-fail path.
• Subagent returns malformed output -> surface what it returned and ask the user whether to retry, fall back to direct (you read frames), or accept as-is.

Bundled scripts

• bin/run.sh cache-aware orchestrator (entry point)
• bin/watch.sh pipeline (download -> transcribe -> optional frames), cache-by-id
• bin/summarize.sh stdin report → claude -p haiku → stdout summary
• bin/download.sh yt-dlp wrapper / local path resolver
• bin/extract-frames.py ffprobe + ffmpeg, tiered fps curve
• bin/transcribe.sh whisperkit-cli wrapper (the swap point)
• bin/transcribe.py WebVTT parser (system python3, no venv)
• video-summary-prompt.md haiku subprocess prompt template
• setup.sh one-time install + model pre-download