/claude-watch — Claude turns a video into study notes

You don't have a video input. This skill gives you one and turns each viewing into a saved notes artifact.

Step 0 — Setup preflight (silent on success)

Run on every /claude-watch invocation:

python3 "${CLAUDE_SKILL_DIR}/scripts/setup.py" --check

Exit codes: 0 ready (silent — proceed), 2 missing binaries, 3 missing API key, 4 both. On non-zero, run the installer:

python3 "${CLAUDE_SKILL_DIR}/scripts/setup.py"

On macOS this auto-brew installs ffmpeg + yt-dlp. On Linux/Windows it prints the right commands. It scaffolds ~/.config/claude-watch/.env (mode 0600) with commented placeholders.

If a Whisper key is still missing afterwards, use AskUserQuestion to ask whether the user has a Groq key (preferred — cheaper, faster) or an OpenAI key, and write it to ~/.config/claude-watch/.env. If they don't want to, run with --no-whisper; videos without native captions will come back frames-only.

When to use

• User pastes a tutorial / lecture / talk URL and asks to study it
• User points at a local screen recording or video and wants notes
• User types /claude-watch <url-or-path> [topic]

How to invoke

Step 1 — parse input. Separate the source (URL or path) from any topic the user mentioned. The topic shapes which sections you emphasize in the notes — pass it through to your synthesis, not to the script.

Step 2 — run the watch script.

python3 "${CLAUDE_SKILL_DIR}/scripts/watch.py" "<source>"

Optional flags:

• --start T / --end T — focus on a section (SS, MM:SS, or HH:MM:SS)
• --max-frames N — lower budget (default 80)
• --resolution W — bump frame width to 1024 px when on-screen text is tiny
• --scene-threshold X — sensitivity (default 0.30; raise for fewer cuts, lower for more)
• --max-gap S — coverage floor in seconds (default 45)
• --whisper groq|openai — force backend
• --no-whisper — disable Whisper entirely
• --out-dir DIR — override library root

Step 3 — read every frame. The script ends with a structured === frames === block listing each frame's path and timestamp. Read them all in parallel — they render as images in your context.

Step 4 — load the transcript. The === transcript === block points to transcript.json (or transcript.window.json for focused mode). Read it — it's a list of {t_start, t_end, text, speaker_break}.

Step 5 — write notes.md to the library directory. Use the strict template below. Save to <library_dir>/notes.md. Then print a 3-line summary to chat:

1. Title and slug
2. Number of sections + key concepts
3. Path to the notes file

Do not delete the library dir. It is the artifact.

Notes template (non-negotiable structure)

# <Video Title>

**Source:** <URL or path>  ·  **Duration:** MM:SS  ·  **Watched:** YYYY-MM-DD

## TLDR
<3-4 sentences: what the video is about and the single most important takeaway.>

## Key Concepts
- **<concept>** — <one-line definition> · `[t=MM:SS]`
- ...

## Notes

### [t=00:04] <Section title you derive from on-screen + spoken content>

![](frames/0001_t00-04.jpg)

**On screen:** <Transcribe / describe the slide, code, diagram. If code, transcribe verbatim.>

**Said:** <Relevant transcript excerpt for this scene, lightly cleaned.>

**Synthesis:** <Your connection — what this section is teaching, how it links to prior section.>

### [t=00:31] <next section>
...

## Code & Commands
<every code-on-screen frame's content as a runnable fenced block, language-tagged, with [t=MM:SS] back-link>

```python
# [t=03:45]
def forward(x):
    return x @ W + b
```

## Diagrams Referenced
- `[t=02:10]` — <one-line description of the diagram in frame 0008>
- ...

## Open Questions
- <things mentioned but not fully covered, or follow-ups to explore>

Rules baked into the template

• One scene = one section. Use the t=MM:SS from each frame as the section anchor.
• Adjacent scenes that are clearly the same topic can be merged. When you do, mention it parenthetically: (merged scenes at t=02:10 and t=02:42)
• Code blocks must be fenced with the right language tag, transcribed verbatim from the frame.
• The "On screen" block is required even for title slides. Keeps the structure parallel.
• Timestamps are absolute (real video timeline) — for YouTube sources, a viewer can paste <URL>&t=<seconds> to jump there.

Re-runs

If the user re-watches the same URL, the script reuses the cached download, transcript, and scenes. Only frames + notes regenerate. To force a full re-run, delete <library_dir>/meta.json first.

Failure modes

• Setup preflight non-zero → run setup.py, then ask for a key via AskUserQuestion.
• No transcript → script emits transcript_source: none. Generate notes frames-only and tell the user.
• Long video sparse-scan warning → offer to re-run with --start/--end focused on the part the user cares about.
• Whisper failure → retry with --whisper openai (if Groq failed) or vice versa.

Token budget

Frames dominate cost (~50-80k input tokens for 60 frames at 512 px). Transcripts are cheap. --resolution 1024 quadruples per-frame cost — only when the user must read tiny on-screen text.

If the user asks a follow-up about a video you already watched in this session, do NOT re-run the script. The library directory is on disk; re-Read only the frames you need.

Security

• Runs yt-dlp, ffmpeg, ffprobe locally
• Sends extracted mono 16 kHz audio to Groq (preferred) or OpenAI Whisper API only when captions are missing
• Reads/writes ~/.config/claude-watch/.env (mode 0600) for keys
• Persists artifacts to ~/claude-watch/library/<slug>/ — review the directory after first run if you're cautious
• Does NOT log or transmit API keys, video files, or the original URL outside the audio-to-Whisper call