lecture

Transcribe a local video lecture using mlx-whisper (Apple Silicon), extract key slide frames with timestamps, then synthesize into a vault-formatted lecture note with embedded screenshots. Auto-detects transcript language — works with any language, outputs English notes.

<Use_When>

• User provides a local video file and wants lecture notes
• User says "take notes from this lecture/video"
• User uses /lecture with a file path
• User has an MP4/MOV/MKV file to process </Use_When>

<Do_Not_Use_When>

• User has a YouTube URL (use /youtube instead)
• User wants to process an existing vault note (use /process)
• User wants audio-only transcription without note synthesis </Do_Not_Use_When>

- `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)

<Execution_Policy>

• Extract first, synthesize second, integrate third
• Always check vault for existing notes on the same topic before creating
• Create note as type: lecture with processing_status: inbox
• The note is a starting point — user can /process it later for deeper engagement
• Transcription can take several minutes for long videos — inform the user </Execution_Policy>

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"

<Tool_Usage>

• Bash: Run extract_lecture.py, copy frames, generate timestamps, search vault
• Read: Read agent definition, read extracted JSON, view frame images for descriptions, read existing vault notes
• Write: Create the lecture note in vault
• Agent: Delegate synthesis to lecture-noter agent (sonnet model)
• Grep/Glob: Search vault for duplicates and related notes </Tool_Usage>

User: /lecture /tmp/risk-management-training.mp4 1. Extract → transcript (48 min video, ~59K chars Vietnamese) + 12 frames with timestamps 2. Copy frames to assets/ as lecture-risk-mgmt-01.jpg through lecture-risk-mgmt-08.jpg 3. Review each frame → build descriptions (title slide, framework diagram, severity table, etc.) 4. Search vault → found 3 related notes on risk and finance topics 5. Agent synthesizes → 6 themed sections, 5 embedded screenshots, 4 questions, 3 concept suggestions 6. Create note → notes/finance/(Lecture) Risk Management Overview.md 7. Report: "Created lecture note with 6 sections and 5 embedded slides. Found connections to [[(Term) Credit Cycle]] and [[(Term) Second-Order Thinking]]. 3 concepts could become Term notes. Run /process when ready." User: /lecture /tmp/risk-management-training.mp4 - Dumps raw transcript into a note without synthesis - Saves screenshots without timestamps, can't trace back to video - Creates chronological summary instead of thematic organization - Misses cross-domain connections - Doesn't review frame content before passing to agent

<Escalation_And_Stop_Conditions>

• uv not installed: Print install command (brew install uv) and stop
• ffmpeg not found: Inform user to install via homebrew
• Transcript error: Report the error, suggest checking audio track
• Video extremely long (>3hrs): Warn user, offer to process first half only
• No audio track: Inform user, offer to extract frames only
• Duplicate note exists: Show existing note, ask if user wants to update or create new </Escalation_And_Stop_Conditions>

$ARGUMENTS