🧠 course-distiller

Turn a course you have access to into a private, LLM-queryable knowledge base, locally, for $0.

Scope: course-distiller is built for Kajabi-style courses (/products/…/categories/…/posts/…) with Wistia video. Other platforms work once someone adds an adapter (see Architecture → Extending). Know this before you install the dependencies.

Great courses are locked inside video players and PDFs. You can't search them, can't ask them questions, and re-watching a 90-minute recording to find one framework is painful. Tools like NotebookLM answer questions over text brilliantly, if only you had the text.

course-distiller produces exactly that: a clean markdown knowledge base: lesson text, distilled video transcripts, extracted PDF/slide content, and every link/contact/form, ready for NotebookLM or any LLM. It runs entirely on your machine. No API bills, no uploading your paid content to a third party.

output/
├── lessons/              # one markdown file per lesson, grouped by section
├── notebooklm/           # consolidated per-section bundles, upload these to NotebookLM
├── _INDEX.md             # clickable map of the whole course
└── PROTOCOLS.md          # auto-aggregated contacts, forms, and resource links

Each lesson file contains the portal text, a distilled transcript of its video, the extracted text of its attachments, and all resource links, verbatim.

It's plain markdown, so it's usable anywhere: drop it into NotebookLM, query it with Claude or ChatGPT, load it into Obsidian, or just grep. NotebookLM is one option, not the destination.

streamlit run app.py

1. Configure. Paste your course's URL and choose where the AI runs: local (free, needs Ollama / Apple-Silicon) or a cloud provider (paste an API key, no GPU required). Keys are passed as environment variables, never written to disk.

   

2. Analyze. A browser opens for you to log in (the tool never sees your password). It crawls the course and shows a pre-flight estimate. Lessons, videos, time, and (for cloud) cost. Nothing expensive runs until you approve it.

3. Proceed. A live dashboard tracks each phase: 🗺️ map → 📄 harvest → 🎬 transcribe → 🔍 extract → 📦 bundle.

4. Download. Grab the markdown and use it anywhere.

Prefer the terminal? cp config.example.yaml config.yaml, edit it, then python -m coursedistiller.run (add --plan to see the estimate first).

flowchart LR
    A[🔑 Your browser<br/>you log in] --> B[🗺️ Crawl<br/>Playwright + fetch]
    B --> C[📄 Parse lessons<br/>HTML→markdown]
    C --> D[🎬 Video<br/>yt-dlp → Whisper]
    D --> E[🧠 Distill<br/>local LLM/Ollama]
    C --> F[📎 Attachments<br/>download + OCR]
    E --> G[🧩 Compile]
    F --> G
    C --> G
    G --> H[📦 NotebookLM<br/>bundles]

See docs/ARCHITECTURE.md for the design decisions (graph-closure crawl, resumability, why the models take turns).

Prerequisites (one-time):

brew install ffmpeg poppler tesseract        # macOS (Linux: apt install ffmpeg poppler-utils tesseract-ocr)
# Ollama for local distillation: https://ollama.com
ollama pull gpt-oss:20b                       # or any local model you like

The tool:

git clone https://github.com/M19K/course-distiller && cd course-distiller
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
playwright install chromium

Whisper (video transcription) is the long pole. Everything else, crawling, parsing, downloading, distilling, is minutes. Transcription time depends almost entirely on your hardware:

• No GPU? Use a cloud provider. Set transcription to Groq and the table above stops mattering, a full course transcribes in minutes for a few cents, and the pre-flight estimate shows the exact number before you commit. This is the "usable by anyone" path.
• Text only (toggle video off): the whole course in a couple of minutes.
• RAM: a 20B distill model needs ~16 GB. On a tight machine, use a smaller Ollama model, or set distill.enabled: false (you'll get raw transcripts instead of notes).
• Tip: use video.max_videos to transcribe a handful first and sanity-check output before a full run.

See examples/ for a complete walkthrough, a sample config.yaml, the command to run, and the exact output tree you should expect (with a real compiled lesson).

Everything lives in config.yaml (copy config.example.yaml): course URL, whether to transcribe video, which Whisper/LLM models, politeness pacing, OCR, and more.

course-distiller is a personal-productivity tool for content you are legitimately entitled to access, like yt-dlp or a read-later app. It automates your own logged-in session.

• Do not redistribute extracted content, it belongs to its creator. The .gitignore ensures no downloaded content is ever committed.
• Do not use it to circumvent access controls or access content you haven't paid for.
• Respect each platform's Terms of Service. You are responsible for how you use this tool.

• Video transcripts in the corpus are distilled (summarized); full verbatim transcripts are kept on disk separately (work/transcripts/).
• Image-only content (diagrams, visual portfolios) isn't OCR'd inside lessons.
• Platform support is Kajabi/Wistia today (see the scope note up top).

• 📐 Architecture: how the pipeline is built and why
• 📝 Product case study, the problem, the decisions, the trade-offs

MIT. See LICENSE.

Built by Maaz Kazi — product and systems, San Francisco.

maazkazi.com · github.com/M19K · linkedin.com/in/maazkazi26