Lecture

• ffmpeg installed (for audio extraction and frame capture)
• uv installed (brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh)
• Apple Silicon Mac (mlx-whisper is optimized for M-series chips)

Stage 1: EXTRACT

Parse the video file path from $ARGUMENTS. If no path provided, ask the user. Verify the file exists and is a video format (mp4, mov, mkv, avi, webm).

Run the extraction script:

SKILL_DIR="${CLAUDE_SKILL_DIR}"
LECTURE_OUTPUT="temp/lecture-extract-output.json"
uv run "$SKILL_DIR/scripts/extract_lecture.py" "VIDEO_PATH" > "$LECTURE_OUTPUT" 2>&1 &

IMPORTANT: This script takes time (several minutes for a 30-60 min video). Inform the user: "Extracting audio and transcribing — this will take a few minutes for a [duration] video."

Run it and wait for completion. Then read the output JSON.

The JSON contains:

• filename, duration, duration_seconds, width, height
• transcript.full_text, transcript.segments (with start/end times), transcript.language
• transcript.error (null if success)
• frames[] — array of {path, timestamp_seconds, timestamp} for each extracted slide
• output_dir — temp directory with extracted frames

If transcript.error is not null: inform the user and stop. Check if mlx-whisper is installed.

If transcript is very long (>80,000 chars): warn the user. Send first 60,000 chars to the agent with a note about total length.

Stage 2: PREPARE FRAMES

Copy the extracted frames to the vault's assets directory with a descriptive naming scheme:

# Generate a slug from the video filename
SLUG=$(echo "VIDEO_FILENAME" | sed 's/\.[^.]*$//' | tr '[:upper:]' '[:lower:]' | tr ' ' '-' | sed 's/[^a-z0-9-]//g' | cut -c1-30)
ASSETS_DIR="assets"

for frame in FRAME_PATHS; do
  FRAME_NUM=$(basename "$frame" | grep -o '[0-9]*')
  cp "$frame" "$ASSETS_DIR/lecture-${SLUG}-${FRAME_NUM}.jpg"
done

Build a frame manifest for the noter agent — each frame gets:

• Its vault filename (for ![[embedding]])
• Its timestamp in the video (e.g., "12:30")

Example manifest:

FRAMES WITH TIMESTAMPS:
- lecture-risk-mgmt-01.jpg (timestamp: 0:10)
- lecture-risk-mgmt-02.jpg (timestamp: 2:00)
- lecture-risk-mgmt-03.jpg (timestamp: 4:00)
...

Stage 3: SYNTHESIZE

Read the agent definition:

Read("${CLAUDE_SKILL_DIR}/agents/lecture-noter.md")

Search the vault for existing notes related to the lecture's topics using the MCP tool:

search_notes(query="KEYWORD", limit=20)

Or fall back to Grep if MCP is unavailable:

Grep(pattern="KEYWORD", path="notes/", glob="*.md", head_limit=20)

Review each frame using the Read tool to see what's on each slide. Build a brief description of each frame's content (1 line each) to include in the agent prompt.

Launch the lecture-noter agent:

Agent(
  subagent_type="general-purpose",
  model="sonnet",
  run_in_background=false,
  prompt="You are Lecture Noter. Follow these instructions exactly:

  [INSERT FULL CONTENT OF agents/lecture-noter.md HERE]

  VIDEO METADATA:
  - Filename: [filename]
  - Duration: [duration]
  - Transcript language: [language from extraction JSON]

  FRAMES WITH TIMESTAMPS AND DESCRIPTIONS:
  - lecture-slug-01.jpg (timestamp: 0:10) — Title slide showing course name
  - lecture-slug-02.jpg (timestamp: 2:00) — Diagram of risk framework
  [... one line per frame with what you see on it]

  EXISTING VAULT NOTES ON RELATED TOPICS:
  [List any matching notes found in grep search]

  TRANSCRIPT:
  [full_text]

  Produce the note body following the Output Format. Do NOT include frontmatter —
  only the body starting from the # title line.
  Use the exact filenames from FRAMES list for ![[embedding]] — do not invent filenames."
)

Stage 4: INTEGRATE

1. Generate timestamp ID:

date +%Y%m%d%H%M%S

2. Determine the best subfolder for the note:

   • ML/AI → notes/ml/
   • Business/startup → notes/startup/
   • Finance → notes/finance/
   • Design → notes/design/
   • Psychology → notes/psychology/
   • General → notes/

3. Create the note file with frontmatter + agent output:

---
id: YYYYMMDDHHMMSS
type: lecture
processing_status: inbox
created_date: YYYY-MM-DD
updated_date: YYYY-MM-DD
---

[AGENT OUTPUT HERE — starts with # title, includes embedded screenshots]

4. Clean up temp files:

rm -rf "$OUTPUT_DIR"
rm -f "$LECTURE_OUTPUT"

5. Report to user:
   • Note path and title
   • Number of screenshots embedded
   • Number of concepts suggested for extraction
   • Any related vault notes found
   • Remind: "Run /process on this note when you're ready to deepen it"

$ARGUMENTS