Video Watch

Video watch

Turn a video into a note a future you can act on without re-watching it. The question this skill answers is "what is the method, is it real, and what does it cost?" — not "what is this video about."

0. Preflight — can Claude see video at all?

Check whether the mcp__claudetube__* tools are in your tool roster (e.g. process_video_tool, get_transcript, get_hq_frames).

• Present → go to §1.

• Absent → stop and set it up. Say plainly: "I can't watch video yet on this machine — it's a one-time setup, about ten minutes." Then follow references/setup.md for the user's OS, one step at a time, checking after each. There is a preflight script that reports exactly what's missing:

  # Mac / Linux
  bash "${CLAUDE_PLUGIN_ROOT}/skills/video-watch/scripts/preflight.sh"
  
  # Windows (PowerShell)
  powershell -ExecutionPolicy Bypass -File "${CLAUDE_PLUGIN_ROOT}\skills\video-watch\scripts\preflight.ps1"
  

  After registering the MCP server the user must restart Claude Code before the tools appear. Tell them that; don't keep retrying in the same session.

Never fake it. If you can't watch the video, do not produce a summary from the title, the URL, the description, or what you "know" about the creator. A confident summary of a video nobody watched is worse than no summary.

1. Ingest

mcp__claudetube__process_video_tool(<url or local file path>)

• First run on a video takes ~30–90 s (download + local transcription). It's cached after that, so follow-up questions are instant.
• Instagram / TikTok URLs sometimes hit a login wall. If ingestion fails, ask the user to download the file (any downloader works) and pass the local path instead. Don't loop on the URL.
• Long videos (>20 min): call get_scenes first and work chapter by chapter. Don't try to hold a 90-minute transcript in your head at once.

2. Triage — decide how much to actually look at

Read the title, uploader, and first ~30 s of transcript, then decide:

Use find_moments_tool("<what you're looking for>") to jump to a specific thing instead of scrubbing. Use ask_video for a quick single question when the user doesn't want the full note.

Watch enough to answer §3 honestly — no more. A 60-second clip needs the transcript and a frame or two. A 20-minute screen-recorded tutorial needs the frames, because the method is on the screen and not in the narration.

3. Write the note

Reply in chat by default. If the user asked you to save it (or has said where notes live), write one file per video: video-notes/<uploader>-<short-slug>.md in the current project, and tell them the path.

# <Title> — <what it claims to do> (<Platform>, <YYYY-MM>)

**Source:** <url> · <duration> · <uploader>
**Watched:** <today's date> via claudetube

## One-line summary
<the method in one sentence — what they do, to get what>

## Shown vs. claimed
<Two short lists. SHOWN = you saw it work on screen or in the transcript with specifics.
 CLAIMED = stated as a result but never demonstrated ("this got me 40 leads a week").
 This section is the whole reason the note exists. Be strict.>

## The method
<ordered steps, in enough detail to rebuild it. If a step was hand-waved, say "(not shown)"
 on that step rather than filling it in yourself.>

## Tools named
<one per line: name — what it's used for — price. Name ONLY what they actually name.
 Price is a real number (subscription $/mo, credits, free-tier limit) or the words
 "cost not stated". Never guess a price.>

## What's missing
<what they skipped, glossed over, or put behind a "link in bio" / lead magnet / paid course>

## Timestamps worth revisiting
<mm:ss — why. Only moments you'd actually want to re-check: a config, a number, a demo.>

## Verdict
**USE / TRY / SKIP** — <one line: why, for THIS user's situation if they've told you it>

Rules that keep the note honest

1. Shown ≠ claimed. A result stated on camera is a claim. A result demonstrated on screen with specifics is shown. Keep them in separate lists, always.
2. Name only what they name. If they say "an AI tool" and never name it, write "an unnamed AI tool" — don't fill in the one you think they meant.
3. Every tool gets a price or "cost not stated." Never invent one.
4. "(not shown)" beats a plausible guess. If a step is hand-waved, mark it. The gap is information.
5. The verdict is one line. If you need a paragraph, you don't have a verdict yet.
6. Don't pad. A thin video gets a thin note. Five honest lines beat forty filler ones.
7. Say what you didn't watch. If you sampled a 2-hour video, state which parts you covered.

4. Follow-ups

After the note, the video stays cached — the user can ask "what did they say about X", "show me the screen at 4:10", "compare this to the other video" without re-ingesting. list_cached_videos shows what's already been watched on this machine.

What this skill is NOT

• Not a transcript dumper. If the user just wants the transcript, get_transcript and hand it over — no note.
• Not a hype filter that always says SKIP. Real methods exist; the point is to separate them from the ones that were only claimed.