Distill

Paste a URL. Pick your level. Get a course.

Distill takes expert-level content — deep blog posts, arXiv papers, GitHub repos — and generates progressive courses with lesson documents, scaffolded exercises, and observable milestones. Each course builds toward reproducing the author's results.

UI_demo.mp4

Lesson documents, not summaries — 3,000–10,000 word teaching documents with running examples, inline code, comprehension checks, and formula walkthroughs.

tuft_demo_small.mp4

Exercises you can actually run — scaffolded .py files with TODO blocks, docstrings, and a __main__ test harness. Solutions included.

vscode_demo.mp4

Install

git clone https://github.com/TAOGenna/Distill.git
cd Distill
uv sync
uv run python -m distill
# → opens http://localhost:8420

How It Works

Phase 1 — Blueprint. Reads the full source material and produces a curriculum: module dependencies, scaffold contracts per exercise (what's provided vs what the student writes), key excerpts, and validation criteria.

Phase 2 — Generate. Each module gets a multi-turn conversation with the full source. The model writes the lesson first (deep processing), then exercises one at a time. Solutions are executed between turns — real output from exercise 1 feeds into exercise 2's prompt. Modules generate in parallel.

Phase 3 — Review. Pre-flight checks (syntax, TODOs, output patterns) catch structural issues. LLM review checks pedagogical quality and contract compliance. Failed modules are re-generated.

Acknowledgments

Inspired by karpathify, Stanford's CS231n assignments, MIT 6.102 course readings, Simon Boehm and Alexa Godric's blog style and beautiful excalidraws.