lecture-processing

process — lecture video → Obsidian notes

Everything needed to turn a recorded lecture into notes: the scripts, a Docker image with the tools, and the agent skill for the LLM steps. The full write-up of how and why is in process.md; the diagram is pipeline.svg.

Layout

process/
├── README.md               this file
├── process.md, pipeline.svg, Lecture processing.pdf    the process, step by step
├── AGENTS.md               vault conventions for coding agents (language, markers, Obsidian links)
├── Dockerfile              tools image; `--target agent` adds Node, Pi and OpenCode
├── docker-run.sh           runs the image with the current folder mounted at /work
├── requirements.txt        Python packages (faster-whisper, scenedetect, imagehash, …)
├── scripts/
│   ├── lecture.sh          pipeline driver: probe | audio | transcribe | proxy | scenes | slides | build | check | all
│   ├── transcribe.py       faster-whisper → "[start → end]  text" lines
│   ├── extract_slides.py   one image per scene from the original video, slides/index.csv
│   ├── build_lecture_md.py transcript + slide index → lecture.md (--obsidian)
│   ├── dump_slides.py      compact, chunked dump of lecture.md for LLM reading
│   └── check_links.py      verifies Obsidian links; exit 1 on broken links
├── skills/lecture-to-notes/
│   ├── SKILL.md            the LLM half: Q & A digest, topic notes, references
│   └── templates/          topics_template.md, qa_digest_template.md
└── legacy/                 one-off scripts from the first lectures (kept for reference)

Quick start (local, with the existing .venv)

source ../.venv/bin/activate
S=process/scripts            # from the vault root

$S/lecture.sh probe -c 1238:804:342:0 Course/lecture.mp4    # check probe.png / check.png
$S/lecture.sh all -c 1238:804:342:0 -e 3:52:37 \
   -p "Лекція про Harness в агентах типу Codex, Claude Code, Pi" \
   -t "Harness Engineering, частина 1" Course/lecture.mp4

Outputs go to Course/lecture/:

• audio.wav
• transcript.txt
• proxy.mp4
• scenes.csv
• slides/
• lecture.md

Steps whose output exists are skipped; -f redoes them. lecture.sh -h lists all options.

Docker

cd process
docker build -t lecture-tools .                          # add --build-arg VAAPI=1 for h264_vaapi on Linux
cd ../Harness_Engineering
../process/docker-run.sh lecture.sh all -c 1238:804:342:0 lecture.mp4

• Whisper models are cached in the lecture-models volume. The first run downloads about 1.6 GB for large-v3-turbo.
• On macOS, Docker has no GPU access. Transcription runs on CPU (int8), and the proxy uses libx264 instead of VideoToolbox. For long lectures, the native .venv on the Mac is faster. The container is for Linux boxes and reproducibility.
• On Linux with an Intel or AMD iGPU, docker-run.sh passes /dev/dri through, and lecture.sh picks h264_vaapi automatically.

Agent layer (optional)

docker build -t lecture-agent --target agent .
IMAGE=lecture-agent ANTHROPIC_API_KEY=… ../process/docker-run.sh pi        # or: opencode

• The image has Node 22, Pi (@mariozechner/pi-coding-agent) and OpenCode (opencode-ai). Change them with --build-arg AGENT_PACKAGES="…".
• The lecture-to-notes skill is installed in ~/.agents/skills/, which both agents read. AGENTS.md is installed as their global instructions.
• API keys are passed from the environment at run time and never stored in the image.
• Claude Code / Cowork: copy or symlink skills/lecture-to-notes into ~/.claude/skills/ (or the vault's .claude/skills/). The same SKILL.md works there.